app.py

import gradio as gr
import cv2
from PIL import Image
import numpy as np
from transformers import pipeline
import os
import torch
import torch.nn.functional as F
from torchvision import transforms
from torchvision.transforms import Compose
#import open3d as o3d
import tempfile
from functools import partial
import spaces
from zipfile import ZipFile
from vincenty import vincenty
import json
#import DracoPy
from collections import Counter
import mediapy

#from depth_anything.dpt import DepthAnything
#from depth_anything.util.transform import Resize, NormalizeImage, PrepareForNet
from huggingface_hub import hf_hub_download
from depth_anything_v2.dpt import DepthAnythingV2

DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
model_configs = {
    'vits': {'encoder': 'vits', 'features': 64, 'out_channels': [48, 96, 192, 384]},
    'vitb': {'encoder': 'vitb', 'features': 128, 'out_channels': [96, 192, 384, 768]},
    'vitl': {'encoder': 'vitl', 'features': 256, 'out_channels': [256, 512, 1024, 1024]},
    'vitg': {'encoder': 'vitg', 'features': 384, 'out_channels': [1536, 1536, 1536, 1536]}
}
encoder2name = {
    'vits': 'Small',
    'vitb': 'Base',
    'vitl': 'Large',
    'vitg': 'Giant', # we are undergoing company review procedures to release our giant model checkpoint
}

edge = []
gradient = None
params = { "fnum":0, "l":16 }
dcolor = []
pcolors = []
frame_selected = 0
frames = []
depths = []
masks = []
locations = []
mesh = []
mesh_n = []
scene = None

def zip_files(files_in, files_out):
    with ZipFile("depth_result.zip", "w") as zipObj:
        for idx, file in enumerate(files_in):
            zipObj.write(file, file.split("/")[-1])
        for idx, file in enumerate(files_out):
            zipObj.write(file, file.split("/")[-1])
    return "depth_result.zip"

def create_video(frames, fps, type):
    print("building video result")
    imgs = []
    for j, img in enumerate(frames):
        imgs.append(cv2.cvtColor(cv2.imread(img).astype(np.uint8), cv2.COLOR_BGR2RGB))

    mediapy.write_video(type + "_result.mp4", imgs, fps=fps)
    return type + "_result.mp4"

@torch.no_grad()
#@spaces.GPU
def predict_depth(image, model):
    return model.infer_image(image)
    
#def predict_depth(model, image):
#    return model(image)["depth"]

def make_video(video_path, outdir='./vis_video_depth', encoder='vits', remove_bg=False):
    if encoder not in ["vitl","vitb","vits"]:
        encoder = "vits"

    model_name = encoder2name[encoder]
    model = DepthAnythingV2(**model_configs[encoder])
    filepath = hf_hub_download(repo_id=f"depth-anything/Depth-Anything-V2-{model_name}", filename=f"depth_anything_v2_{encoder}.pth", repo_type="model")
    state_dict = torch.load(filepath, map_location="cpu")
    model.load_state_dict(state_dict)
    model = model.to(DEVICE).eval()

    #mapper = {"vits":"small","vitb":"base","vitl":"large"}
    # DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
    # model = DepthAnything.from_pretrained('LiheYoung/depth_anything_vitl14').to(DEVICE).eval()
    # Define path for temporary processed frames
    #temp_frame_dir = tempfile.mkdtemp()
    
    #margin_width = 50
    #to_tensor_transform = transforms.ToTensor()

    #DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
    # depth_anything = DepthAnything.from_pretrained('LiheYoung/depth_anything_{}14'.format(encoder)).to(DEVICE).eval()
    #depth_anything = pipeline(task = "depth-estimation", model=f"nielsr/depth-anything-{mapper[encoder]}")
    
    # total_params = sum(param.numel() for param in depth_anything.parameters())
    # print('Total parameters: {:.2f}M'.format(total_params / 1e6))
    
    #transform = Compose([
    #    Resize(
    #        width=518,
    #        height=518,
    #        resize_target=False,
    #        keep_aspect_ratio=True,
    #        ensure_multiple_of=14,
    #        resize_method='lower_bound',
    #        image_interpolation_method=cv2.INTER_CUBIC,
    #    ),
    #    NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    #    PrepareForNet(),
    #])

    if os.path.isfile(video_path):
        if video_path.endswith('txt'):
            with open(video_path, 'r') as f:
                lines = f.read().splitlines()
        else:
            filenames = [video_path]
    else:
        filenames = os.listdir(video_path)
        filenames = [os.path.join(video_path, filename) for filename in filenames if not filename.startswith('.')]
        filenames.sort()
    
    # os.makedirs(outdir, exist_ok=True)
    
    for k, filename in enumerate(filenames):
        file_size = os.path.getsize(filename)/1024/1024
        if file_size > 128.0:
            print(f'File size of {filename} larger than 128Mb, sorry!')
            return filename
        print('Progress {:}/{:},'.format(k+1, len(filenames)), 'Processing', filename)
        
        raw_video = cv2.VideoCapture(filename)
        frame_width, frame_height = int(raw_video.get(cv2.CAP_PROP_FRAME_WIDTH)), int(raw_video.get(cv2.CAP_PROP_FRAME_HEIGHT))
        frame_rate = int(raw_video.get(cv2.CAP_PROP_FPS))
        if frame_rate < 1:
            frame_rate = 1
        cframes = int(raw_video.get(cv2.CAP_PROP_FRAME_COUNT))
        print(f'frames: {cframes}, fps: {frame_rate}')
        # output_width = frame_width * 2 + margin_width
        
        #filename = os.path.basename(filename)
        # output_path = os.path.join(outdir, filename[:filename.rfind('.')] + '_video_depth.mp4')
        #with tempfile.NamedTemporaryFile(delete=False, suffix='.mp4') as tmpfile:
        #    output_path = tmpfile.name
        #out = cv2.VideoWriter(output_path, cv2.VideoWriter_fourcc(*"avc1"), frame_rate, (output_width, frame_height))
        #fourcc = cv2.VideoWriter_fourcc(*'mp4v')
        #out = cv2.VideoWriter(output_path, fourcc, frame_rate, (output_width, frame_height))
        global masks
        count = 0
        n = 0
        depth_frames = []
        orig_frames = []
        thumbnail_old = []

        while raw_video.isOpened():
            ret, raw_frame = raw_video.read()
            if not ret:
                break
            else:
                print(count)

            frame = cv2.cvtColor(raw_frame, cv2.COLOR_BGR2RGB) / 255.0
            frame_pil = Image.fromarray((frame * 255).astype(np.uint8))
            #frame = transform({'image': frame})['image']
            #frame = torch.from_numpy(frame).unsqueeze(0).to(DEVICE)

            #
            depth = predict_depth(raw_frame[:, :, ::-1], model)
            depth = (depth - depth.min()) / (depth.max() - depth.min()) * 255.0
            depth = depth.astype(np.uint8)
            depth_color = Image.fromarray(depth)
            depth_gray = cv2.cvtColor(depth_color, cv2.COLOR_RGBA2GRAY)
            #
            
            #depth = to_tensor_transform(predict_depth(depth_anything, frame_pil))
            #depth = F.interpolate(depth[None], (frame_height, frame_width), mode='bilinear', align_corners=False)[0, 0]
            #depth = (depth - depth.min()) / (depth.max() - depth.min()) * 255.0
            #depth = depth.cpu().numpy().astype(np.uint8)
            #depth_color = cv2.applyColorMap(depth, cv2.COLORMAP_BONE)
            #depth_gray = cv2.cvtColor(depth_color, cv2.COLOR_RGBA2GRAY)

            # Remove white border around map:
            # define lower and upper limits of white
            #white_lo = np.array([250,250,250])
            #white_hi = np.array([255,255,255])
            # mask image to only select white
            mask = cv2.inRange(depth_gray[0:int(depth_gray.shape[0]/8*6.5)-1, 0:depth_gray.shape[1]], 250, 255)
            # change image to black where we found white
            depth_gray[0:int(depth_gray.shape[0]/8*6.5)-1, 0:depth_gray.shape[1]][mask>0] = 0

            mask = cv2.inRange(depth_gray[int(depth_gray.shape[0]/8*6.5):depth_gray.shape[0], 0:depth_gray.shape[1]], 160, 255)
            depth_gray[int(depth_gray.shape[0]/8*6.5):depth_gray.shape[0], 0:depth_gray.shape[1]][mask>0] = 160

            depth_color = cv2.cvtColor(depth_gray, cv2.COLOR_GRAY2BGR)
            # split_region = np.ones((frame_height, margin_width, 3), dtype=np.uint8) * 255
            # combined_frame = cv2.hconcat([raw_frame, split_region, depth_color])
            
            # out.write(combined_frame)
            # frame_path = os.path.join(temp_frame_dir, f"frame_{count:05d}.png")
            # cv2.imwrite(frame_path, combined_frame)

            #raw_frame = cv2.cvtColor(raw_frame, cv2.COLOR_BGR2BGRA)
            #raw_frame[:, :, 3] = 255

            if remove_bg == True:
                if count >= 1: #int(cframes/2):
                    n = 0 #n = count-int(cframes/2)
                    depth_color_bg = cv2.imread(f"f{n}_dmap.png").astype(np.uint8)
                    raw_frame_bg = cv2.imread(f"f{n}.png").astype(np.uint8)
                    
                    diff_d = np.abs(depth_color.astype(np.int16)-depth_color_bg.astype(np.int16))
                    diff_c = np.abs(raw_frame.astype(np.int16)-raw_frame_bg.astype(np.int16))
                    avg_d = int(np.average(diff_d))
                    avg_c = int(np.average(diff_c))
                    print('-average')
                    print(avg_d)
                    print(avg_c)
                    md_d = int(np.median(diff_d))
                    md_c = int(np.median(diff_c))
                    print('-median')
                    print(md_d)
                    print(md_c)
                    print('-most common')
                    c = Counter(diff_d.flatten())
                    value, cc = c.most_common()[0]
                    print(value)
                    print(cc)
                    c = Counter(diff_c.flatten())
                    value, cc = c.most_common()[0]
                    print(value)
                    print(cc)
                    print('-')
                    
                    mask_bg = cv2.inRange(diff_d, np.array([0,0,0]), np.array([md_d,md_d,md_d]))
                    mask_no_shadow = cv2.inRange(diff_c, np.array([0,0,0]), np.array([md_c,md_c,md_c]))
                    mask_shadow = cv2.bitwise_not(mask_no_shadow)

                    #stereo = cv2.StereoBM.create(numDisparities=16, blockSize=15)
                    #disparity = stereo.compute(raw_frame_l, raw_frame_r)
                    
                    m = cv2.inRange(raw_frame, np.array([255,255,255]), np.array([255,255,255]))
                    raw_frame[m>0] = (239,239,239)
                    raw_frame[cv2.bitwise_and(mask_shadow, mask_bg)>0] = (raw_frame[cv2.bitwise_and(mask_shadow, mask_bg)>0] / 17 + 240).astype(np.uint8)
                    raw_frame[cv2.bitwise_and(mask_no_shadow, mask_bg)>0] = (255,255,255)
            
            thumbnail = cv2.cvtColor(cv2.resize(raw_frame, (16,32)), cv2.COLOR_BGR2GRAY).flatten()
            if len(thumbnail_old) > 0:
                diff = thumbnail - thumbnail_old
                #print(diff)
                c = Counter(diff)
                value, cc = c.most_common()[0]
                if value == 0 and cc > int(16*32*0.8):
                    continue
            thumbnail_old = thumbnail
            
            cv2.imwrite(f"f{count}.png", raw_frame)
            orig_frames.append(f"f{count}.png")
            
            cv2.imwrite(f"f{count}_dmap.png", depth_color)
            depth_frames.append(f"f{count}_dmap.png")

            cv2.imwrite(f"f{count}_mask.png", depth_gray)
            masks.append(f"f{count}_mask.png")
            count += 1

        if remove_bg == True:
            final_vid = create_video(orig_frames, frame_rate, "orig")
        else:
            final_vid = create_video(depth_frames, frame_rate, "depth")
            
        final_zip = zip_files(orig_frames, depth_frames)
        raw_video.release()
        # out.release()
        cv2.destroyAllWindows()

        global gradient
        global frame_selected
        global depths
        global frames
        frames = orig_frames
        depths = depth_frames

        if depth_color.shape[0] == 2048: #height
            gradient = cv2.imread('./gradient_large.png').astype(np.uint8)
        elif depth_color.shape[0] == 1024:
            gradient = cv2.imread('./gradient.png').astype(np.uint8)
        else:
            gradient = cv2.imread('./gradient_small.png').astype(np.uint8)
        
        return final_vid, final_zip, frames, masks[frame_selected], depths #output_path

def depth_edges_mask(depth):
    """Returns a mask of edges in the depth map.
    Args:
    depth: 2D numpy array of shape (H, W) with dtype float32.
    Returns:
    mask: 2D numpy array of shape (H, W) with dtype bool.
    """
    # Compute the x and y gradients of the depth map.
    depth_dx, depth_dy = np.gradient(depth)
    # Compute the gradient magnitude.
    depth_grad = np.sqrt(depth_dx ** 2 + depth_dy ** 2)
    # Compute the edge mask.
    mask = depth_grad > 0.05
    return mask

def pano_depth_to_world_points(depth):
    """
    360 depth to world points
    given 2D depth is an equirectangular projection of a spherical image
    Treat depth as radius
    longitude : -pi to pi
    latitude : -pi/2 to pi/2
    """

    # Convert depth to radius
    radius = (255 - depth.flatten())

    lon = np.linspace(0, np.pi*2, depth.shape[1])
    lat = np.linspace(0, np.pi, depth.shape[0])
    lon, lat = np.meshgrid(lon, lat)
    lon = lon.flatten()
    lat = lat.flatten()

    pts3d = [[0,0,0]]
    uv = [[0,0]]
    nl = [[0,0,0]]
    for i in range(0, 1): #(0,2)
        for j in range(0, 1): #(0,2)
            #rnd_lon = (np.random.rand(depth.shape[0]*depth.shape[1]) - 0.5) / 8
            #rnd_lat = (np.random.rand(depth.shape[0]*depth.shape[1]) - 0.5) / 8
            d_lon = lon + i/2 * np.pi*2 / depth.shape[1]
            d_lat = lat + j/2 * np.pi / depth.shape[0]

            nx = np.cos(d_lon) * np.sin(d_lat)
            ny = np.cos(d_lat)
            nz = np.sin(d_lon) * np.sin(d_lat)
            
            # Convert to cartesian coordinates
            x = radius * nx
            y = radius * ny
            z = radius * nz

            pts = np.stack([x, y, z], axis=1)
            uvs = np.stack([lon/np.pi/2, lat/np.pi], axis=1)
            nls = np.stack([-nx, -ny, -nz], axis=1)
            
            pts3d = np.concatenate((pts3d, pts), axis=0)
            uv = np.concatenate((uv, uvs), axis=0)
            nl = np.concatenate((nl, nls), axis=0)
            #print(f'i: {i}, j: {j}')
            j = j+1
        i = i+1
        
    return [pts3d, uv, nl]

def rgb2gray(rgb):
    return np.dot(rgb[...,:3], [0.333, 0.333, 0.333])

def get_mesh(image, depth, blur_data, loadall):
    global depths
    global pcolors
    global frame_selected
    global mesh
    global mesh_n
    global scene
    if loadall == False:
        mesh = []
        mesh_n = []
    fnum = frame_selected

    #print(image[fnum][0])
    #print(depth["composite"])

    depthc = cv2.imread(depths[frame_selected], cv2.IMREAD_UNCHANGED).astype(np.uint8)
    blur_img = blur_image(cv2.imread(image[fnum][0], cv2.IMREAD_UNCHANGED).astype(np.uint8), depthc, blur_data)
    gdepth = cv2.cvtColor(depthc, cv2.COLOR_RGB2GRAY) #rgb2gray(depthc)
    
    print('depth to gray - ok')
    points = pano_depth_to_world_points(gdepth)
    pts3d = points[0]
    uv = points[1]
    nl = points[2]
    print('radius from depth - ok')

    # Create a trimesh mesh from the points
    # Each pixel is connected to its 4 neighbors
    # colors are the RGB values of the image
    uvs = uv.reshape(-1, 2)
    #print(uvs)
    #verts = pts3d.reshape(-1, 3)
    verts = [[0,0,0]]
    normals = nl.reshape(-1, 3)
    rgba = cv2.cvtColor(blur_img, cv2.COLOR_RGB2RGBA)
    colors = rgba.reshape(-1, 4)
    clrs = [[128,128,128,0]]

    #for i in range(0,1): #(0,4)
    #clrs = np.concatenate((clrs, colors), axis=0)
        #i = i+1
    #verts, clrs

    #pcd = o3d.geometry.TriangleMesh.create_tetrahedron()
    #pcd.compute_vertex_normals()
    #pcd.paint_uniform_color((1.0, 1.0, 1.0))
    #mesh.append(pcd)
    #print(mesh[len(mesh)-1])
    if not str(fnum) in mesh_n:
        mesh_n.append(str(fnum))
    print('mesh - ok')

    # Save as glb
    glb_file = tempfile.NamedTemporaryFile(suffix='.glb', delete=False)
    #o3d.io.write_triangle_mesh(glb_file.name, pcd)
    print('file - ok')
    return "./TriangleWithoutIndices.gltf", glb_file.name, ",".join(mesh_n)

def blur_image(image, depth, blur_data):
    blur_a = blur_data.split()
    print(f'blur data {blur_data}')

    blur_frame = image.copy()
    j = 0
    while j < 256:
        i = 255 - j
        blur_lo = np.array([i,i,i])
        blur_hi = np.array([i+1,i+1,i+1])
        blur_mask = cv2.inRange(depth, blur_lo, blur_hi)
        
        print(f'kernel size {int(blur_a[j])}')
        blur = cv2.GaussianBlur(image, (int(blur_a[j]), int(blur_a[j])), 0)
                
        blur_frame[blur_mask>0] = blur[blur_mask>0]
        j = j + 1

    return blur_frame

def loadfile(f):
    return f

def show_json(txt):
    data = json.loads(txt)
    print(txt)
    i=0
    while i < len(data[2]):
        data[2][i] = data[2][i]["image"]["path"]
        data[4][i] = data[4][i]["path"]
        i=i+1
    return data[0]["video"]["path"], data[1]["path"], data[2], data[3]["background"]["path"], data[4], data[5]


def select_frame(d, evt: gr.SelectData):
    global dcolor
    global frame_selected
    global masks
    global edge
    
    if evt.index != frame_selected:
        edge = []
        mask = cv2.imread(depths[frame_selected]).astype(np.uint8)
        cv2.imwrite(masks[frame_selected], cv2.cvtColor(mask, cv2.COLOR_RGB2GRAY))
        frame_selected = evt.index

    if len(dcolor) == 0:
        bg = [127, 127, 127, 255]
    else:
        bg = "[" + str(dcolor[frame_selected])[1:-1] + ", 255]"
        
    return masks[frame_selected], frame_selected, bg

def switch_rows(v):
    global frames
    global depths
    if v == True:
        print(depths[0])
        return depths
    else:
        print(frames[0])
        return frames

def optimize(v, d):
    global pcolors
    global dcolor
    global frame_selected
    global frames
    global depths
    
    if v == True:
        ddepth = cv2.CV_16S
        kernel_size = 3
        l = 16

        dcolor = []
        for k, f in enumerate(frames):
            frame = cv2.imread(frames[k]).astype(np.uint8)
            
            # convert to np.float32
            f = np.float32(frame.reshape((-1,3)))
            # define criteria, number of clusters(K) and apply kmeans()
            criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 4, 1.0)
            ret,label,center=cv2.kmeans(f,l,None,criteria,4,cv2.KMEANS_RANDOM_CENTERS)
            # Now convert back into uint8, and make original image
            center = np.uint8(center)
            res = center[label.flatten()]
            frame = res.reshape((frame.shape))

            depth = cv2.imread(depths[k]).astype(np.uint8)
            mask = cv2.cvtColor(depth, cv2.COLOR_RGB2GRAY)
            dcolor.append(bincount(frame[mask==0]))
            print(dcolor[k])
            clrs = Image.fromarray(frame.astype(np.uint8)).convert('RGB').getcolors()
            i=0
            while i<len(clrs):
                clrs[i] = list(clrs[i][1])
                clrs[i].append(255)
                i=i+1
            print(clrs)
            pcolors = clrs
            
            #mask = cv2.convertScaleAbs(cv2.Laplacian(cv2.cvtColor(frame, cv2.COLOR_RGB2GRAY), ddepth, ksize=kernel_size))
            #mask[mask>0] = 255
            #frame[mask==0] = (0, 0, 0)
            cv2.imwrite(frames[k], frame)

            #depth[mask==0] = (255,255,255)
            mask = cv2.inRange(frame, np.array([dcolor[k][0]-8, dcolor[k][1]-8, dcolor[k][2]-8]), np.array([dcolor[k][0]+8, dcolor[k][1]+8, dcolor[k][2]+8]))
            depth[mask>0] = (255,255,255)
            depth[depth.shape[0]-1:depth.shape[0], 0:depth.shape[1]] = (160, 160, 160)
            depth[0:1, 0:depth.shape[1]] = (0, 0, 0)
            cv2.imwrite(depths[k], depth)
            
    if d == False:
      return frames, "[" + str(dcolor[frame_selected])[1:-1] + ", 255]"
    else:
      return depths, "[" + str(dcolor[frame_selected])[1:-1] + ", 255]"

def bincount(a):
    a2D = a.reshape(-1,a.shape[-1])
    col_range = (256, 256, 256) # generically : a2D.max(0)+1
    a1D = np.ravel_multi_index(a2D.T, col_range)
    return list(reversed(np.unravel_index(np.bincount(a1D).argmax(), col_range)))

def reset_mask():
    global frame_selected
    global masks
    global depths
    global edge

    edge = []
    mask = cv2.imread(depths[frame_selected]).astype(np.uint8)
    cv2.imwrite(masks[frame_selected], cv2.cvtColor(mask, cv2.COLOR_RGB2GRAY))
    return masks[frame_selected], depths

def apply_mask(d, b):
    global frames
    global frame_selected
    global masks
    global depths
    global edge

    edge = []
    mask = cv2.cvtColor(d["layers"][0], cv2.COLOR_RGBA2GRAY)
    mask[mask<255] = 0
    b = b*2+1
    dilation = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2 * b + 1, 2 * b + 1), (b, b))
    mask = cv2.dilate(mask, dilation)
    mask_b = cv2.GaussianBlur(mask, (b,b), 0)
    b = b*2+1
    dilation = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2 * b + 1, 2 * b + 1), (b, b))
    dmask = cv2.dilate(mask, dilation)
    dmask_b = cv2.GaussianBlur(dmask, (b,b), 0)

    for k, mk in enumerate(masks):
        if k != frame_selected and k < len(depths):
            cv2.imwrite(masks[k], dmask)
            frame = cv2.imread(frames[k], cv2.IMREAD_UNCHANGED).astype(np.uint8)
            frame[:, :, 3] = dmask_b
            cv2.imwrite(frames[k], frame)
        
    frame = cv2.imread(frames[frame_selected], cv2.IMREAD_UNCHANGED).astype(np.uint8)
    frame[:, :, 3] = 255 - mask_b
    cv2.imwrite(frames[frame_selected], frame)
    
    cv2.imwrite(masks[frame_selected], mask) #d["background"]
    return masks[frame_selected], depths, frames

def draw_mask(l, t, v, d, evt: gr.EventData):
    global depths
    global params
    global frame_selected
    global masks
    global gradient
    global edge
    
    points = json.loads(v)
    pts = np.array(points, np.int32)
    pts = pts.reshape((-1,1,2))

    if len(edge) == 0 or params["fnum"] != frame_selected or params["l"] != l:
      if len(edge) > 0:
          d["background"] = cv2.imread(depths[frame_selected]).astype(np.uint8)
          
          if d["background"].shape[0] == 2048: #height
            gradient = cv2.imread('./gradient_large.png').astype(np.uint8)
          elif d["background"].shape[0] == 1024:
            gradient = cv2.imread('./gradient.png').astype(np.uint8)
          else:
            gradient = cv2.imread('./gradient_small.png').astype(np.uint8)

      bg = cv2.cvtColor(d["background"], cv2.COLOR_RGBA2GRAY)

      diff = np.abs(bg.astype(np.int16)-cv2.cvtColor(gradient, cv2.COLOR_RGBA2GRAY).astype(np.int16)).astype(np.uint8)
      mask = cv2.inRange(diff, 0, t)
      #kernel = np.ones((c,c),np.float32)/(c*c)
      #mask = cv2.filter2D(mask,-1,kernel)
      dilation = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (15-(t*2+1), 15-(t*2+1)), (t, t))
      mask = cv2.dilate(mask, dilation)

      #indices = np.arange(0,256)   # List of all colors 
      #divider = np.linspace(0,255,l+1)[1] # we get a divider
      #quantiz = np.intp(np.linspace(0,255,l)) # we get quantization colors
      #color_levels = np.clip(np.intp(indices/divider),0,l-1) # color levels 0,1,2..
      #palette = quantiz[color_levels]
      
      #for i in range(l):
      #    bg[(bg >= i*255/l) & (bg < (i+1)*255/l)] = i*255/(l-1)
      #bg = cv2.convertScaleAbs(palette[bg]).astype(np.uint8) # Converting image back to uint

      res = np.float32(bg.reshape((-1,1)))
      criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 4, 1.0)
      ret,label,center=cv2.kmeans(res,l,None,criteria,4,cv2.KMEANS_PP_CENTERS)
      center = np.uint8(center)
      res = center[label.flatten()]
      bg = res.reshape((bg.shape))
        
      bg[mask>0] = 0
      bg[bg==255] = 0

      params["fnum"] = frame_selected
      params["l"] = l

      d["layers"][0] = cv2.cvtColor(bg, cv2.COLOR_GRAY2RGBA)
      edge = bg.copy()
    else:
      bg = edge.copy()

    x = points[len(points)-1][0]
    y = points[len(points)-1][1]

                                                #int(t*256/l)
    mask = cv2.floodFill(bg, None, (x, y), 1, 0, 256, (4 | cv2.FLOODFILL_FIXED_RANGE))[2] #(4 | cv2.FLOODFILL_FIXED_RANGE | cv2.FLOODFILL_MASK_ONLY | 255 << 8)
    # 255 << 8 tells to fill with the value 255)
    mask = mask[1:mask.shape[0]-1, 1:mask.shape[1]-1]
    
    d["layers"][0][mask>0] = (255,255,255,255)
    
    return gr.ImageEditor(value=d)


def findNormals(format):
    global depths
    d_im = cv2.cvtColor(cv2.imread(depths[frame_selected]).astype(np.uint8), cv2.COLOR_BGR2GRAY)
    zy, zx = np.gradient(d_im)  
    # You may also consider using Sobel to get a joint Gaussian smoothing and differentation
    # to reduce noise
    #zx = cv2.Sobel(d_im, cv2.CV_64F, 1, 0, ksize=5)     
    #zy = cv2.Sobel(d_im, cv2.CV_64F, 0, 1, ksize=5)

    if format == "opengl":
        zy = -zy
        
    normal = np.dstack((np.ones_like(d_im), -zy, -zx))
    n = np.linalg.norm(normal, axis=2)
    normal[:, :, 0] /= n
    normal[:, :, 1] /= n
    normal[:, :, 2] /= n

    # offset and rescale values to be in 0-255
    normal += 1
    normal /= 2
    normal *= 255

    return (normal[:, :, ::-1]).astype(np.uint8)


load_model="""
async(c, o, b, p, d, n, m)=>{
  var intv = setInterval(function(){
    if (document.getElementById("iframe3D")===null || typeof document.getElementById("iframe3D")==="undefined") {
      try {
      if (typeof BABYLON !== "undefined" && BABYLON.Engine && BABYLON.Engine.LastCreatedScene) {
        BABYLON.Engine.LastCreatedScene.onAfterRenderObservable.add(function() { //onDataLoadedObservable

          var then = new Date().getTime();
          var now, delta;
          const interval = 1000 / 25;
          const tolerance = 0.1;
          BABYLON.Engine.LastCreatedScene.getEngine().stopRenderLoop();
          BABYLON.Engine.LastCreatedScene.getEngine().runRenderLoop(function () {
            now = new Date().getTime();
            delta = now - then;
            then = now - (delta % interval);
            if (delta >= interval - tolerance) {
                BABYLON.Engine.LastCreatedScene.render();
            }
          });
          
          var bg = JSON.parse(document.getElementById("bgcolor").getElementsByTagName("textarea")[0].value);
          BABYLON.Engine.LastCreatedScene.getEngine().setHardwareScalingLevel(1.0);
          for (var i=0; i<bg.length; i++) {
            bg[i] /= 255;
          }
          BABYLON.Engine.LastCreatedScene.clearColor = new BABYLON.Color4(bg[0], bg[1], bg[2], bg[3]);
          BABYLON.Engine.LastCreatedScene.ambientColor = new BABYLON.Color4(255,255,255,255);
          //BABYLON.Engine.LastCreatedScene.autoClear = false;
          //BABYLON.Engine.LastCreatedScene.autoClearDepthAndStencil = false;
          for (var i=0; i<BABYLON.Engine.LastCreatedScene.getNodes().length; i++) {
            if (BABYLON.Engine.LastCreatedScene.getNodes()[i].material) {
              BABYLON.Engine.LastCreatedScene.getNodes()[i].material.pointSize = Math.ceil(Math.log2(Math.PI/document.getElementById("zoom").value));
            }
          }
          BABYLON.Engine.LastCreatedScene.getAnimationRatio();
          //BABYLON.Engine.LastCreatedScene.activeCamera.inertia = 0.0;
        });
        if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
          BABYLON.Engine.LastCreatedScene.activeCamera.metadata = {
            pipeline: new BABYLON.DefaultRenderingPipeline("default", true, BABYLON.Engine.LastCreatedScene, [BABYLON.Engine.LastCreatedScene.activeCamera]) 
          }
        }
        BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.samples = 4;
        BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.contrast = 1.0;
        BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.exposure = 1.0;

        BABYLON.Engine.LastCreatedScene.activeCamera.fov = document.getElementById("zoom").value;

        document.getElementById("model3D").getElementsByTagName("canvas")[0].style.filter = "blur(" + Math.ceil(Math.log2(Math.PI/document.getElementById("zoom").value))/2.0*Math.sqrt(2.0) + "px)";
        document.getElementById("model3D").getElementsByTagName("canvas")[0].oncontextmenu = function(e){e.preventDefault();}
        document.getElementById("model3D").getElementsByTagName("canvas")[0].ondrag = function(e){e.preventDefault();}

        if (o.indexOf(""+n) < 0) {
          if (o != "") { o += ","; }
          o += n;
        }
        //alert(o);
        var o_ = o.split(",");
        var q = BABYLON.Engine.LastCreatedScene.meshes;
        for(i = 0; i < q.length; i++) {
          let mesh = q[i];
          mesh.dispose(false, true);
        }
        var dome = [];
        for (var j=0; j<o_.length; j++) {
          o_[j] = parseInt(o_[j]);
          dome[j] = new BABYLON.PhotoDome("dome"+j, p[o_[j]].image.url, 
          {
            resolution: 16,
            size: 512
          }, BABYLON.Engine.LastCreatedScene);
          var q = BABYLON.Engine.LastCreatedScene.meshes[BABYLON.Engine.LastCreatedScene.meshes.length-2]._children;
          for(i = 0; i < q.length; i++) {
            let mesh = q[i];
            mesh.dispose(false, true);
          }
          //BABYLON.Engine.LastCreatedScene.meshes[BABYLON.Engine.LastCreatedScene.meshes.length-1].material.needDepthPrePass = true;
          //BABYLON.Engine.LastCreatedScene.meshes[BABYLON.Engine.LastCreatedScene.meshes.length-1].scaling.z = -1;
          BABYLON.Engine.LastCreatedScene.meshes[BABYLON.Engine.LastCreatedScene.meshes.length-1].alphaIndex = o_.length-j;
          BABYLON.Engine.LastCreatedScene.meshes[BABYLON.Engine.LastCreatedScene.meshes.length-1].material.diffuseTexture.hasAlpha = true;
          BABYLON.Engine.LastCreatedScene.meshes[BABYLON.Engine.LastCreatedScene.meshes.length-1].material.useAlphaFromDiffuseTexture = true;
          BABYLON.Engine.LastCreatedScene.meshes[BABYLON.Engine.LastCreatedScene.meshes.length-1].applyDisplacementMap(m[o_[j]].url, 0, 255, function(m){try{alert(BABYLON.Engine.Version);}catch(e){alert(e);}}, null, null, true, function(e){alert(e);});
        }
        clearInterval(intv);
      }
      } catch(e) {alert(e);}
    } else if (BABYLON || BABYLON == null) {
    try {
      BABYLON = null;
      if (document.getElementById("model3D").getElementsByTagName("canvas")[0]) {
        document.getElementById("model3D").getElementsByTagName("canvas")[0].remove();
      }
      document.getElementById("iframe3D").src = "index.htm";
      document.getElementById("iframe3D").onload = function() {
        if (o.indexOf(""+n) < 0) {
          if (o != "") { o += ","; }
          o += n;
        }
        alert(o);
        var o_ = o.split(",");
        document.getElementById("iframe3D").contentDocument.getElementById("coords").value = c;
        document.getElementById("iframe3D").contentDocument.getElementById("order").value = o;
        document.getElementById("iframe3D").contentDocument.getElementById("bgcolor").value = b;
        document.getElementById("iframe3D").contentDocument.getElementById("bgimage").value = "";
        document.getElementById("iframe3D").contentDocument.getElementById("bgdepth").value = "";
        for (var j=0; j<o_.length; j++) {
          o_[j] = parseInt(o_[j]);
          alert(o_[j]);
          document.getElementById("iframe3D").contentDocument.getElementById("bgimage").value += p[o_[j]].image.url + ",";
          document.getElementById("iframe3D").contentDocument.getElementById("bgdepth").value += m[o_[j]].url + ",";
        }
      }
      toggleDisplay("model");
      
      clearInterval(intv);
    } catch(e) {alert(e)}
    }
  }, 40);
}
"""

js = """
async()=>{
  console.log('Hi');

const chart = document.getElementById('chart');
const blur_in = document.getElementById('blur_in').getElementsByTagName('textarea')[0];
var md = false;
var xold = 128;
var yold = 32;
var a = new Array(256);
var l;

for (var i=0; i<256; i++) {
  const hr = document.createElement('hr');
  hr.style.backgroundColor = 'hsl(0,0%,' + (100-i/256*100) + '%)';
  chart.appendChild(hr);
}

function resetLine() {
  a.fill(1);
  for (var i=0; i<256; i++) {
    chart.childNodes[i].style.height = a[i] + 'px';
    chart.childNodes[i].style.marginTop = '32px';
  }
}
resetLine();
window.resetLine = resetLine;

function pointerDown(x, y) {
  md = true;
  xold = parseInt(x - chart.getBoundingClientRect().x);
  yold = parseInt(y - chart.getBoundingClientRect().y);
  chart.title = xold + ',' + yold;
}
window.pointerDown = pointerDown;

function pointerUp() {
  md = false;
  var evt = document.createEvent('Event');
  evt.initEvent('input', true, false);
  blur_in.dispatchEvent(evt);
  chart.title = '';
}
window.pointerUp = pointerUp;

function lerp(y1, y2, mu) { return y1*(1-mu)+y2*mu; }

function drawLine(x, y) {
  x = parseInt(x - chart.getBoundingClientRect().x);
  y = parseInt(y - chart.getBoundingClientRect().y);
  if (md === true && y >= 0 && y < 64 && x >= 0 && x < 256) {
    if (y < 32) {
      a[x] = Math.abs(32-y)*2 + 1;
      chart.childNodes[x].style.height = a[x] + 'px';
      chart.childNodes[x].style.marginTop = y + 'px';

      for (var i=Math.min(xold, x)+1; i<Math.max(xold, x); i++) {
        l = parseInt(lerp( yold, y, (i-xold)/(x-xold) ));

        if (l < 32) {
          a[i] = Math.abs(32-l)*2 + 1;
          chart.childNodes[i].style.height = a[i] + 'px';
          chart.childNodes[i].style.marginTop = l + 'px';
        } else if (l < 64) {
          a[i] = Math.abs(l-32)*2 + 1;
          chart.childNodes[i].style.height = a[i] + 'px';
          chart.childNodes[i].style.marginTop = (64-l) + 'px';
        }
      }
    } else if (y < 64) {
      a[x] = Math.abs(y-32)*2 + 1;
      chart.childNodes[x].style.height = a[x] + 'px';
      chart.childNodes[x].style.marginTop = (64-y) + 'px';

      for (var i=Math.min(xold, x)+1; i<Math.max(xold, x); i++) {
        l = parseInt(lerp( yold, y, (i-xold)/(x-xold) ));

        if (l < 32) {
          a[i] = Math.abs(32-l)*2 + 1;
          chart.childNodes[i].style.height = a[i] + 'px';
          chart.childNodes[i].style.marginTop = l + 'px';
        } else if (l < 64) {
          a[i] = Math.abs(l-32)*2 + 1;
          chart.childNodes[i].style.height = a[i] + 'px';
          chart.childNodes[i].style.marginTop = (64-l) + 'px';
        }
      }
    }
    blur_in.value = a.join(' ');
    xold = x;
    yold = y;
    chart.title = xold + ',' + yold;
  }
}
window.drawLine = drawLine;
  
}
"""

css = """
#img-display-container {
    max-height: 100vh;
    }
#img-display-input {
    max-height: 80vh;
    }
#img-display-output {
    max-height: 80vh;
    }
"""

title = "# Depth Anything V2 Video"
description = """**Depth Anything V2** on full video files.
Please refer to our [paper](https://arxiv.org/abs/2406.09414), [project page](https://depth-anything-v2.github.io), and [github](https://github.com/DepthAnything/Depth-Anything-V2) for more details."""

    
#transform = Compose([
#        Resize(
#            width=518,
#            height=518,
#            resize_target=False,
#            keep_aspect_ratio=True,
#            ensure_multiple_of=14,
#            resize_method='lower_bound',
#            image_interpolation_method=cv2.INTER_CUBIC,
#        ),
#        NormalizeImage(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
#        PrepareForNet(),
#])

# @torch.no_grad()
# def predict_depth(model, image):
#     return model(image)

with gr.Blocks(css=css, js=js) as demo:
    gr.Markdown(title)
    gr.Markdown(description)
    gr.Markdown("### Video Depth Prediction demo")

    with gr.Row():
        with gr.Column():
            input_json = gr.Textbox(elem_id="json_in", value="{}", label="JSON", interactive=False)
            input_url = gr.Textbox(elem_id="url_in", value="./examples/streetview.mp4", label="URL")
            input_video = gr.Video(label="Input Video", format="mp4")
            input_url.input(fn=loadfile, inputs=[input_url], outputs=[input_video])
            submit = gr.Button("Submit")
            output_frame = gr.Gallery(label="Frames", preview=True, columns=8192, interactive=False)
            output_switch = gr.Checkbox(label="Show depths")
            output_depth = gr.Files(label="Depths", interactive=False)
            output_switch.input(fn=switch_rows, inputs=[output_switch], outputs=[output_frame])
            optimize_switch = gr.Checkbox(label="Optimize")
            bgcolor = gr.Textbox(elem_id="bgcolor", value="[127, 127, 127, 255]", label="Background color", interactive=False)
            optimize_switch.input(fn=optimize, inputs=[optimize_switch, output_switch], outputs=[output_frame, bgcolor])
            output_mask = gr.ImageEditor(layers=False, sources=('upload', 'clipboard'), show_download_button=True, type="numpy", interactive=True, transforms=(None,), eraser=gr.Eraser(), brush=gr.Brush(default_size=0, colors=['black', '#505050', '#a0a0a0', 'white']), elem_id="image_edit")
            with gr.Row():
              selector = gr.HTML(value="""
            <a href='#' id='selector' onclick='if (this.style.fontWeight!=\"bold\") {
            this.style.fontWeight=\"bold\";
            document.getElementById(\"image_edit\").getElementsByTagName(\"canvas\")[0].oncontextmenu = function(e){e.preventDefault();}
            document.getElementById(\"image_edit\").getElementsByTagName(\"canvas\")[0].ondrag = function(e){e.preventDefault();}
            
            document.getElementById(\"image_edit\").getElementsByTagName(\"canvas\")[0].onclick = function(e) {
              var x = parseInt((e.clientX-e.target.getBoundingClientRect().x)*e.target.width/e.target.getBoundingClientRect().width);
              var y = parseInt((e.clientY-e.target.getBoundingClientRect().y)*e.target.height/e.target.getBoundingClientRect().height);

              var p = document.getElementById(\"mouse\").getElementsByTagName(\"textarea\")[0].value.slice(1, -1);
              if (p != \"\") { p += \", \"; }
              p += \"[\" + x + \", \" + y + \"]\";
              document.getElementById(\"mouse\").getElementsByTagName(\"textarea\")[0].value = \"[\" + p + \"]\";
              
              var evt = document.createEvent(\"Event\");
              evt.initEvent(\"input\", true, false);
              document.getElementById(\"mouse\").getElementsByTagName(\"textarea\")[0].dispatchEvent(evt);
            }
            document.getElementById(\"image_edit\").getElementsByTagName(\"canvas\")[0].onpointerdown = function(e) {
              
              document.getElementById(\"mouse\").getElementsByTagName(\"textarea\")[0].style.borderColor = \"#a0a0a0\";
            
            }
            document.getElementById(\"image_edit\").getElementsByTagName(\"canvas\")[0].onpointerup = function(e) {
              
              document.getElementById(\"mouse\").getElementsByTagName(\"textarea\")[0].style.borderColor = \"#ffffff\";
            
            }
            } else {
              this.style.fontWeight=\"normal\";
              document.getElementById(\"image_edit\").getElementsByTagName(\"canvas\")[0].onclick = null;
              
            }' title='Select point' style='text-decoration:none;color:white;'>⊹ Select point</a> <a href='#' id='clear_select' onclick='
              
              document.getElementById(\"mouse\").getElementsByTagName(\"textarea\")[0].value = \"[]\";
            
            ' title='Clear selection' style='text-decoration:none;color:white;'>✕ Clear</a>""")
              apply = gr.Button("Apply", size='sm')
              reset = gr.Button("Reset", size='sm')
            with gr.Accordion(label="Edge", open=False):
              levels = gr.Slider(label="Color levels", value=16, maximum=32, minimum=2, step=1)
              tolerance = gr.Slider(label="Tolerance", value=1, maximum=7, minimum=0, step=1)
              bsize = gr.Slider(label="Border size", value=15, maximum=256, minimum=1, step=2)
              mouse = gr.Textbox(elem_id="mouse", value="""[]""", interactive=False)
              mouse.input(fn=draw_mask, show_progress="minimal", inputs=[levels, tolerance, mouse, output_mask], outputs=[output_mask])
              apply.click(fn=apply_mask, inputs=[output_mask, bsize], outputs=[output_mask, output_depth, output_frame])
              reset.click(fn=reset_mask, inputs=None, outputs=[output_mask, output_depth])

            normals_out = gr.Image(label="Normal map", interactive=False)
            format_normals = gr.Radio(choices=["directx", "opengl"])
            find_normals = gr.Button("Find normals")
            find_normals.click(fn=findNormals, inputs=[format_normals], outputs=[normals_out])

        with gr.Column():
            model_type = gr.Dropdown([("small", "vits"), ("base", "vitb"), ("large", "vitl")], type="value", value="vits", label='Model Type')
            remove_bg = gr.Checkbox(label="Remove background")
            processed_video = gr.Video(label="Output Video", format="mp4", interactive=False)
            processed_zip = gr.File(label="Output Archive", interactive=False)
            result = gr.Model3D(label="3D Mesh", clear_color=[0.5, 0.5, 0.5, 0.0], camera_position=[0, 90, 0], zoom_speed=2.0, pan_speed=2.0, interactive=True, elem_id="model3D") #, display_mode="point_cloud"
            chart_c = gr.HTML(elem_id="chart_c", value="""<div id='chart' onpointermove='window.drawLine(event.clientX, event.clientY);' onpointerdown='window.pointerDown(event.clientX, event.clientY);' onpointerup='window.pointerUp();' onpointerleave='window.pointerUp();' onpointercancel='window.pointerUp();' onclick='window.resetLine();'></div>
            <style>
  body {
    user-select: none;
  }
  #chart hr {
    width: 1px;
    height: 1px;
    clear: none;
    border: 0;
    padding:0;
    display: inline-block;
    position: relative;
    vertical-align: top;
    margin-top:32px;
  }
  #chart {
    padding:0;
    margin:0;
    width:256px;
    height:64px;
    background-color:#808080;
    touch-action: none;
  }
            </style>
            """)
            average = gr.HTML(value="""<label for='average'>Average</label><input id='average' type='range' style='width:256px;height:1em;' value='1' min='1' max='15' step='2' onclick='
              var pts_a = document.getElementById(\"blur_in\").getElementsByTagName(\"textarea\")[0].value.split(\" \");
              for (var i=0; i<256; i++) {
                var avg = 0;
                var div = this.value;
                for (var j = i-parseInt(this.value/2); j <= i+parseInt(this.value/2); j++) {
                  if (pts_a[j]) {
                    avg += parseInt(pts_a[j]);
                  } else if (div > 1) {
                    div--;
                  }
                }
                pts_a[i] = Math.round((avg / div - 1) / 2) * 2 + 1;

                document.getElementById(\"chart\").childNodes[i].style.height = pts_a[i] + \"px\";
                document.getElementById(\"chart\").childNodes[i].style.marginTop = (64-pts_a[i])/2 + \"px\";
              }
              document.getElementById(\"blur_in\").getElementsByTagName(\"textarea\")[0].value = pts_a.join(\" \");

              var evt = document.createEvent(\"Event\");
              evt.initEvent(\"input\", true, false);
              document.getElementById(\"blur_in\").getElementsByTagName(\"textarea\")[0].dispatchEvent(evt);
            ' oninput='
              this.parentNode.childNodes[2].innerText = this.value;
            ' onchange='this.click();'/><span>1</span>""")
            with gr.Accordion(label="Blur levels", open=False):
                blur_in = gr.Textbox(elem_id="blur_in", label="Kernel size", show_label=False, interactive=False, value="1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1")
            with gr.Accordion(label="Locations", open=False):
                offset = gr.HTML(value="""<input type='text' id='kbrd' onpointerdown='this.style.color = \"white\";' onpointerup='this.style.color = \"auto\";' onpointermove='
                try {
                if (this.style.color!=\"auto\" && BABYLON) {
                  if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
                    var evt = document.createEvent(\"Event\");
                    evt.initEvent(\"click\", true, false);
                    document.getElementById(\"reset_cam\").dispatchEvent(evt);
                  }
                  event.preventDefault();
                  if (BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotationQuaternion) {
                    BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotationQuaternion = null;
                  }
                  const dir = BABYLON.Engine.LastCreatedScene.activeCamera.getForwardRay().direction;
                  dir.y = 0; dir.normalize();
                  const angle = BABYLON.Vector3.GetAngleBetweenVectors(dir, BABYLON.Vector3.Forward(), BABYLON.Vector3.Up());
                  const x = event.clientX-this.getBoundingClientRect().x-128;
                  const y = event.clientY-this.getBoundingClientRect().y-64;
                  const angle_ = Math.atan2(y, x);
                  const r = Math.sqrt(Math.pow(y,2) + Math.pow(x,2));
                  BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.z = r * Math.sin(-angle_-angle);
                  BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.x = r * Math.cos(-angle_-angle);
                }
                } catch(e) {alert(e)}
                ' onkeydown='
                if (BABYLON) {
                  if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
                    var evt = document.createEvent(\"Event\");
                    evt.initEvent(\"click\", true, false);
                    document.getElementById(\"reset_cam\").dispatchEvent(evt);
                  }
                  event.preventDefault();
                  if (BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotationQuaternion) {
                    BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotationQuaternion = null;
                  }
                  try {
                  const dir = BABYLON.Engine.LastCreatedScene.activeCamera.getForwardRay().direction;
                  dir.y = 0; dir.normalize();
                  const angle = BABYLON.Vector3.GetAngleBetweenVectors(dir, BABYLON.Vector3.Forward(), BABYLON.Vector3.Up());
                  switch(event.key) {
                    case \"w\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.y += 1;
                      this.value = \"w ⬆ x\";
                      break;
                    case \"x\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.y -= 1;
                      this.value = \"w ⬇ x\";
                      break;
                    case \"a\":
                      const x = -1; const y = 0;
                      const angle_ = Math.atan2(y, x);
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.z += Math.sin(-angle_-angle);
                      this.value = \"a ⬅ d\";
                      break;
                    case \"d\":
                      const x = 1; const y = 0;
                      const angle_ = Math.atan2(y, x);
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.z += Math.sin(-angle_-angle);
                      this.value = \"a ➡ d\";
                      break;
                    case \"e\":
                      const x = 0; const y = -1;
                      const angle_ = Math.atan2(y, x);
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.x += Math.cos(-angle_-angle);
                      this.value = \"z ↗ e\";
                      break;
                    case \"z\":
                      const x = 0; const y = 1;
                      const angle_ = Math.atan2(y, x);
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.x += Math.cos(-angle_-angle);
                      this.value = \"z ↙ e\";
                      break;
                    case \"s\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.x = 0;
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.y = 0;
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].position.z = 0;
                      this.value = \"\";
                      break;
                    case \"t\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.z += Math.PI/256;
                      this.value = \"t 🔃 b\";
                      break;
                    case \"b\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.z -= Math.PI/256;
                      this.value = \"t 🔃 b\";
                      break;
                    case \"f\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.y -= Math.PI/256;
                      this.value = \"f 🔁 h\";
                      break;
                    case \"h\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.y += Math.PI/256;
                      this.value = \"f 🔁 h\";
                      break;
                    case \"y\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.x -= Math.PI/256;
                      this.value = \"v 🔄 y\";
                      break;
                    case \"v\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.x += Math.PI/256;
                      this.value = \"v 🔄 y\";
                      break;
                    case \"g\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.x = 0;
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.y = 0;
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].rotation.z = 0;
                      this.value = \"\";
                      break;
                    case \"i\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.y *= 256/255;
                      this.value = \"i ↕ ,\";
                      break;
                    case \",\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.y /= 256/255;
                      this.value = \"i ↕ ,\";
                      break;
                    case \"j\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.z /= 256/255;
                      this.value = \"j ↔ l\";
                      break;
                    case \"l\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.z *= 256/255;
                      this.value = \"j ↔ l\";
                      break;
                    case \"o\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.x /= 256/255;
                      this.value = \"m ⤢ o\";
                      break;
                    case \"m\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.x *= 256/255;
                      this.value = \"m ⤢ o\";
                      break;
                    case \"k\":
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.x = 1;
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.y = 1;
                      BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].scaling.z = 1;
                      this.value = \"\";
                      break;
                    default:
                      this.value = \"\"; 
                  }
                  } catch(e) {alert(e)}
                }
                ' style='height:128px;width:256px;user-select:none;touch-action:none;color:auto;background-color:transparent;border:1px solid gray;'/>
                <pre id='keymap'>
`  1  2  3  4  5  6  7  8  9  0  -  =  
       W  E     T  Y     I  O     {  }
     A-`S´-D  F-`G´-H  J-`K´-L  ;  '
      Z´ X̀     V´ B̀     M´ `,  .  /
      <a id='move' href='#'>move</a>    <a id='rotate' href='#'>rotate</a>    <a id='scale' href='#'>scale</a>
                </pre>""")
                selected = gr.Number(elem_id="fnum", value=0, minimum=0, maximum=256, interactive=False)
                output_frame.select(fn=select_frame, inputs=[output_mask], outputs=[output_mask, selected, bgcolor])
                example_coords = """[
                  {"lat": 50.07379596793083, "lng": 14.437146122950555, "heading": 152.70303, "pitch": 2.607833999999997}, 
                  {"lat": 50.073799567020004, "lng": 14.437146774240507, "heading": 151.12973, "pitch": 2.8672300000000064}, 
                  {"lat": 50.07377647505558, "lng": 14.437161000659017, "heading": 151.41025, "pitch": 3.4802200000000028}, 
                  {"lat": 50.07379496839027, "lng": 14.437148958238538, "heading": 151.93391, "pitch": 2.843050000000005}, 
                  {"lat": 50.073823157821664, "lng": 14.437124189538856, "heading": 152.95769, "pitch": 4.233024999999998}
                ]"""
                coords = gr.Textbox(elem_id="coords", value=example_coords, label="Coordinates", interactive=False)
                mesh_order = gr.Textbox(elem_id="order", value="", label="Order", interactive=False)
                
            result_file = gr.File(elem_id="file3D", label="3D file", interactive=False)
            html = gr.HTML(value="""<label for='zoom'>Zoom</label><input id='zoom' type='range' style='width:256px;height:1em;' value='0.8' min='0.157' max='1.57' step='0.001' oninput='
              if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
                var evt = document.createEvent(\"Event\");
                evt.initEvent(\"click\", true, false);
                document.getElementById(\"reset_cam\").dispatchEvent(evt);
              } 
              BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].material.pointSize = Math.ceil(Math.log2(Math.PI/this.value));
              BABYLON.Engine.LastCreatedScene.activeCamera.fov = this.value;
              this.parentNode.childNodes[2].innerText = BABYLON.Engine.LastCreatedScene.activeCamera.fov;

              document.getElementById(\"model3D\").getElementsByTagName(\"canvas\")[0].style.filter = \"blur(\" + BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].material.pointSize/2.0*Math.sqrt(2.0) + \"px)\";
            '/><span>0.8</span>""")
            camera = gr.HTML(value="""<a href='#' id='reset_cam' onclick='
              if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
                BABYLON.Engine.LastCreatedScene.activeCamera.metadata = { 
                  screenshot: true,
                  pipeline: new BABYLON.DefaultRenderingPipeline(\"default\", true, BABYLON.Engine.LastCreatedScene, [BABYLON.Engine.LastCreatedScene.activeCamera]) 
                }
              } 
              BABYLON.Engine.LastCreatedScene.activeCamera.radius = 0;
              BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1].material.pointSize = Math.ceil(Math.log2(Math.PI/document.getElementById(\"zoom\").value));
              BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.samples = 4; 
              BABYLON.Engine.LastCreatedScene.activeCamera.fov = document.getElementById(\"zoom\").value;
              BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.contrast = document.getElementById(\"contrast\").value;
              BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.exposure = document.getElementById(\"exposure\").value;
              
              document.getElementById(\"model3D\").getElementsByTagName(\"canvas\")[0].style.filter = \"blur(\" + Math.ceil(Math.log2(Math.PI/document.getElementById(\"zoom\").value))/2.0*Math.sqrt(2.0) + \"px)\";
              document.getElementById(\"model3D\").getElementsByTagName(\"canvas\")[0].oncontextmenu = function(e){e.preventDefault();}
              document.getElementById(\"model3D\").getElementsByTagName(\"canvas\")[0].ondrag = function(e){e.preventDefault();}

              if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager) {
                BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager = new BABYLON.GizmoManager(BABYLON.Engine.LastCreatedScene, 16);
                
                BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.positionGizmoEnabled = true;
                BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.rotationGizmoEnabled = false;
                BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.scaleGizmoEnabled = false;
                BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.boundingBoxGizmoEnabled = false;
              
                BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.usePointerToAttachGizmos = false;
                document.getElementById(\"move\").onclick = function(event) {
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.positionGizmoEnabled = true;
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.rotationGizmoEnabled = false;
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.scaleGizmoEnabled = false;
                }
                document.getElementById(\"rotate\").onclick = function(event) {
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.positionGizmoEnabled = false;
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.rotationGizmoEnabled = true;
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.scaleGizmoEnabled = false;
                }
                document.getElementById(\"scale\").onclick = function(event) {
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.positionGizmoEnabled = false;
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.rotationGizmoEnabled = false;
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.scaleGizmoEnabled = true;
                }
              }
              BABYLON.Engine.LastCreatedScene.activeCamera.metadata.gizmoManager.attachToMesh(BABYLON.Engine.LastCreatedScene.getNodes()[parseInt(document.getElementById(\"fnum\").getElementsByTagName(\"input\")[0].value)+1]);
            '>reset camera</a>""")
            contrast = gr.HTML(value="""<label for='contrast'>Contrast</label><input id='contrast' type='range' style='width:256px;height:1em;' value='1.0' min='0' max='2' step='0.001' oninput='
              if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
                var evt = document.createEvent(\"Event\");
                evt.initEvent(\"click\", true, false);
                document.getElementById(\"reset_cam\").dispatchEvent(evt);
              } 
              BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.contrast = this.value;
              this.parentNode.childNodes[2].innerText = BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.contrast;
            '/><span>1.0</span>""")
            exposure = gr.HTML(value="""<label for='exposure'>Exposure</label><input id='exposure' type='range' style='width:256px;height:1em;' value='1.0' min='0' max='2' step='0.001' oninput='
              if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
                var evt = document.createEvent(\"Event\");
                evt.initEvent(\"click\", true, false);
                document.getElementById(\"reset_cam\").dispatchEvent(evt);
              } 
              BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.exposure = this.value;
              this.parentNode.childNodes[2].innerText = BABYLON.Engine.LastCreatedScene.activeCamera.metadata.pipeline.imageProcessing.exposure;
            '/><span>1.0</span>""")
            canvas = gr.HTML(value="""<a href='#' onclick='
              if (!BABYLON.Engine.LastCreatedScene.activeCamera.metadata) {
                var evt = document.createEvent(\"Event\");
                evt.initEvent(\"click\", true, false);
                document.getElementById(\"reset_cam\").dispatchEvent(evt);
              } 
              BABYLON.Engine.LastCreatedScene.activeCamera.metadata.screenshot = true;

              BABYLON.Engine.LastCreatedScene.getEngine().onEndFrameObservable.add(function() {
                if (BABYLON.Engine.LastCreatedScene.activeCamera.metadata.screenshot === true) {
                  BABYLON.Engine.LastCreatedScene.activeCamera.metadata.screenshot = false;
                  try {
                    BABYLON.Tools.CreateScreenshotUsingRenderTarget(BABYLON.Engine.LastCreatedScene.getEngine(), BABYLON.Engine.LastCreatedScene.activeCamera, 
                      { precision: 1.0 }, (durl) => { 
                        var cnvs = document.getElementById(\"model3D\").getElementsByTagName(\"canvas\")[0]; //.getContext(\"webgl2\");
                        var svgd = `<svg id=\"svg_out\" viewBox=\"0 0 ` + cnvs.width + ` ` + cnvs.height + `\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\">
                          <defs>
                            <filter id=\"blur\" x=\"0\" y=\"0\" xmlns=\"http://www.w3.org/2000/svg\">
                              <feGaussianBlur in=\"SourceGraphic\" stdDeviation=\"` + BABYLON.Engine.LastCreatedScene.getNodes()[1].material.pointSize/2.0*Math.sqrt(2.0) + `\" />
                            </filter>
                          </defs>
                          <image filter=\"url(#blur)\" id=\"svg_img\" x=\"0\" y=\"0\" width=\"` + cnvs.width + `\" height=\"` + cnvs.height + `\" xlink:href=\"` + durl + `\"/>
                        </svg>`;
                        document.getElementById(\"cnv_out\").width = cnvs.width;
                        document.getElementById(\"cnv_out\").height = cnvs.height;
                        document.getElementById(\"img_out\").src = \"data:image/svg+xml;base64,\" + btoa(svgd);        
                      }
                    );
                  } catch(e) { alert(e); }
                  // https://forum.babylonjs.com/t/best-way-to-save-to-jpeg-snapshots-of-scene/17663/11
                }
              });
            '/>snapshot</a><br/><img src='' id='img_out' onload='
              var ctxt = document.getElementById(\"cnv_out\").getContext(\"2d\");
              ctxt.drawImage(this, 0, 0); 
            '/><br/>
            <canvas id='cnv_out'/>""")
            load_all = gr.Checkbox(label="Load all")
            render = gr.Button("Render")
            input_json.input(show_json, inputs=[input_json], outputs=[processed_video, processed_zip, output_frame, output_mask, output_depth, coords])
    
    def on_submit(uploaded_video,model_type,remove_bg,coordinates):
        global locations
        locations = []
        avg = [0, 0]
        
        locations = json.loads(coordinates)
        for k, location in enumerate(locations):
            if "tiles" in locations[k]:
                locations[k]["heading"] = locations[k]["tiles"]["originHeading"]
                locations[k]["pitch"] = locations[k]["tiles"]["originPitch"]
            else:
                locations[k]["heading"] = 0
                locations[k]["pitch"] = 0

            if "location" in locations[k]:
                locations[k] = locations[k]["location"]["latLng"]
                avg[0] = avg[0] + locations[k]["lat"]
                avg[1] = avg[1] + locations[k]["lng"]
            else:
                locations[k]["lat"] = 0
                locations[k]["lng"] = 0
                
        if len(locations) > 0:
            avg[0] = avg[0] / len(locations)
            avg[1] = avg[1] / len(locations)
            
        for k, location in enumerate(locations):
            lat = vincenty((location["lat"], 0), (avg[0], 0)) * 1000
            lng = vincenty((0, location["lng"]), (0, avg[1])) * 1000
            locations[k]["lat"] = float(lat / 2.5 * 95 * np.sign(location["lat"]-avg[0]))
            locations[k]["lng"] = float(lng / 2.5 * 95 * np.sign(location["lng"]-avg[1]))
        print(locations)
            
        # Process the video and get the path of the output video
        output_video_path = make_video(uploaded_video,encoder=model_type,remove_bg=remove_bg)

        return output_video_path + (json.dumps(locations),)

    submit.click(on_submit, inputs=[input_video, model_type, remove_bg, coords], outputs=[processed_video, processed_zip, output_frame, output_mask, output_depth, coords])
    render.click(None, inputs=[coords, mesh_order, bgcolor, output_frame, output_mask, selected, output_depth], outputs=None, js=load_model)
    render.click(partial(get_mesh), inputs=[output_frame, output_mask, blur_in, load_all], outputs=[result, result_file, mesh_order])

    example_files = [["./examples/streetview.mp4", "vits", False, example_coords], ["./examples/man-in-museum-reverse.mp4", "vits", True, example_coords]]
    examples = gr.Examples(examples=example_files, fn=on_submit, cache_examples=True, inputs=[input_video, model_type, remove_bg, coords], outputs=[processed_video, processed_zip, output_frame, output_mask, output_depth, coords])
    

if __name__ == '__main__':
    demo.queue().launch()