initial commit
Browse files
app.py
ADDED
@@ -0,0 +1,289 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
# gradio for visual demo
|
4 |
+
import gradio as gr
|
5 |
+
|
6 |
+
# transformers for easy access to nnet
|
7 |
+
os.system("pip install git+https://github.com/huggingface/transformers.git")
|
8 |
+
os.system("pip install datasets")
|
9 |
+
os.system("pip install scipy")
|
10 |
+
os.system("pip install torch")
|
11 |
+
|
12 |
+
from transformers import AutoImageProcessor, AutoModelForImageClassification, DPTForDepthEstimation, Mask2FormerForUniversalSegmentation
|
13 |
+
import torch
|
14 |
+
import numpy as np
|
15 |
+
from PIL import Image
|
16 |
+
import requests
|
17 |
+
from collections import defaultdict
|
18 |
+
|
19 |
+
palette = np.asarray([
|
20 |
+
[0, 0, 0],
|
21 |
+
[120, 120, 120],
|
22 |
+
[180, 120, 120],
|
23 |
+
[6, 230, 230],
|
24 |
+
[80, 50, 50],
|
25 |
+
[4, 200, 3],
|
26 |
+
[120, 120, 80],
|
27 |
+
[140, 140, 140],
|
28 |
+
[204, 5, 255],
|
29 |
+
[230, 230, 230],
|
30 |
+
[4, 250, 7],
|
31 |
+
[224, 5, 255],
|
32 |
+
[235, 255, 7],
|
33 |
+
[150, 5, 61],
|
34 |
+
[120, 120, 70],
|
35 |
+
[8, 255, 51],
|
36 |
+
[255, 6, 82],
|
37 |
+
[143, 255, 140],
|
38 |
+
[204, 255, 4],
|
39 |
+
[255, 51, 7],
|
40 |
+
[204, 70, 3],
|
41 |
+
[0, 102, 200],
|
42 |
+
[61, 230, 250],
|
43 |
+
[255, 6, 51],
|
44 |
+
[11, 102, 255],
|
45 |
+
[255, 7, 71],
|
46 |
+
[255, 9, 224],
|
47 |
+
[9, 7, 230],
|
48 |
+
[220, 220, 220],
|
49 |
+
[255, 9, 92],
|
50 |
+
[112, 9, 255],
|
51 |
+
[8, 255, 214],
|
52 |
+
[7, 255, 224],
|
53 |
+
[255, 184, 6],
|
54 |
+
[10, 255, 71],
|
55 |
+
[255, 41, 10],
|
56 |
+
[7, 255, 255],
|
57 |
+
[224, 255, 8],
|
58 |
+
[102, 8, 255],
|
59 |
+
[255, 61, 6],
|
60 |
+
[255, 194, 7],
|
61 |
+
[255, 122, 8],
|
62 |
+
[0, 255, 20],
|
63 |
+
[255, 8, 41],
|
64 |
+
[255, 5, 153],
|
65 |
+
[6, 51, 255],
|
66 |
+
[235, 12, 255],
|
67 |
+
[160, 150, 20],
|
68 |
+
[0, 163, 255],
|
69 |
+
[140, 140, 140],
|
70 |
+
[250, 10, 15],
|
71 |
+
[20, 255, 0],
|
72 |
+
[31, 255, 0],
|
73 |
+
[255, 31, 0],
|
74 |
+
[255, 224, 0],
|
75 |
+
[153, 255, 0],
|
76 |
+
[0, 0, 255],
|
77 |
+
[255, 71, 0],
|
78 |
+
[0, 235, 255],
|
79 |
+
[0, 173, 255],
|
80 |
+
[31, 0, 255],
|
81 |
+
[11, 200, 200],
|
82 |
+
[255, 82, 0],
|
83 |
+
[0, 255, 245],
|
84 |
+
[0, 61, 255],
|
85 |
+
[0, 255, 112],
|
86 |
+
[0, 255, 133],
|
87 |
+
[255, 0, 0],
|
88 |
+
[255, 163, 0],
|
89 |
+
[255, 102, 0],
|
90 |
+
[194, 255, 0],
|
91 |
+
[0, 143, 255],
|
92 |
+
[51, 255, 0],
|
93 |
+
[0, 82, 255],
|
94 |
+
[0, 255, 41],
|
95 |
+
[0, 255, 173],
|
96 |
+
[10, 0, 255],
|
97 |
+
[173, 255, 0],
|
98 |
+
[0, 255, 153],
|
99 |
+
[255, 92, 0],
|
100 |
+
[255, 0, 255],
|
101 |
+
[255, 0, 245],
|
102 |
+
[255, 0, 102],
|
103 |
+
[255, 173, 0],
|
104 |
+
[255, 0, 20],
|
105 |
+
[255, 184, 184],
|
106 |
+
[0, 31, 255],
|
107 |
+
[0, 255, 61],
|
108 |
+
[0, 71, 255],
|
109 |
+
[255, 0, 204],
|
110 |
+
[0, 255, 194],
|
111 |
+
[0, 255, 82],
|
112 |
+
[0, 10, 255],
|
113 |
+
[0, 112, 255],
|
114 |
+
[51, 0, 255],
|
115 |
+
[0, 194, 255],
|
116 |
+
[0, 122, 255],
|
117 |
+
[0, 255, 163],
|
118 |
+
[255, 153, 0],
|
119 |
+
[0, 255, 10],
|
120 |
+
[255, 112, 0],
|
121 |
+
[143, 255, 0],
|
122 |
+
[82, 0, 255],
|
123 |
+
[163, 255, 0],
|
124 |
+
[255, 235, 0],
|
125 |
+
[8, 184, 170],
|
126 |
+
[133, 0, 255],
|
127 |
+
[0, 255, 92],
|
128 |
+
[184, 0, 255],
|
129 |
+
[255, 0, 31],
|
130 |
+
[0, 184, 255],
|
131 |
+
[0, 214, 255],
|
132 |
+
[255, 0, 112],
|
133 |
+
[92, 255, 0],
|
134 |
+
[0, 224, 255],
|
135 |
+
[112, 224, 255],
|
136 |
+
[70, 184, 160],
|
137 |
+
[163, 0, 255],
|
138 |
+
[153, 0, 255],
|
139 |
+
[71, 255, 0],
|
140 |
+
[255, 0, 163],
|
141 |
+
[255, 204, 0],
|
142 |
+
[255, 0, 143],
|
143 |
+
[0, 255, 235],
|
144 |
+
[133, 255, 0],
|
145 |
+
[255, 0, 235],
|
146 |
+
[245, 0, 255],
|
147 |
+
[255, 0, 122],
|
148 |
+
[255, 245, 0],
|
149 |
+
[10, 190, 212],
|
150 |
+
[214, 255, 0],
|
151 |
+
[0, 204, 255],
|
152 |
+
[20, 0, 255],
|
153 |
+
[255, 255, 0],
|
154 |
+
[0, 153, 255],
|
155 |
+
[0, 41, 255],
|
156 |
+
[0, 255, 204],
|
157 |
+
[41, 0, 255],
|
158 |
+
[41, 255, 0],
|
159 |
+
[173, 0, 255],
|
160 |
+
[0, 245, 255],
|
161 |
+
[71, 0, 255],
|
162 |
+
[122, 0, 255],
|
163 |
+
[0, 255, 184],
|
164 |
+
[0, 92, 255],
|
165 |
+
[184, 255, 0],
|
166 |
+
[0, 133, 255],
|
167 |
+
[255, 214, 0],
|
168 |
+
[25, 194, 194],
|
169 |
+
[102, 255, 0],
|
170 |
+
[92, 0, 255],
|
171 |
+
])
|
172 |
+
|
173 |
+
depth_image_processor = AutoImageProcessor.from_pretrained("facebook/dpt-dinov2-small-nyu")
|
174 |
+
depth_model = DPTForDepthEstimation.from_pretrained("facebook/dpt-dinov2-small-nyu")
|
175 |
+
|
176 |
+
def compute_depth(img):
|
177 |
+
# prepare image for the model
|
178 |
+
inputs = depth_image_processor(images=img, return_tensors="pt")
|
179 |
+
|
180 |
+
with torch.no_grad():
|
181 |
+
outputs = depth_model(**inputs)
|
182 |
+
predicted_depth = outputs.predicted_depth
|
183 |
+
|
184 |
+
# interpolate to original size
|
185 |
+
prediction = torch.nn.functional.interpolate(
|
186 |
+
predicted_depth.unsqueeze(1),
|
187 |
+
size=img.size[::-1],
|
188 |
+
mode="bicubic",
|
189 |
+
align_corners=False,
|
190 |
+
)
|
191 |
+
|
192 |
+
# visualize the prediction
|
193 |
+
output = prediction.squeeze().cpu().numpy()
|
194 |
+
formatted = (output * 255 / np.max(output)).astype("uint8")
|
195 |
+
depth = Image.fromarray(formatted)
|
196 |
+
return [depth, "depth"]
|
197 |
+
|
198 |
+
clas_processor = AutoImageProcessor.from_pretrained('facebook/dinov2-small-imagenet1k-1-layer')
|
199 |
+
clas_model = AutoModelForImageClassification.from_pretrained('facebook/dinov2-small-imagenet1k-1-layer')
|
200 |
+
|
201 |
+
def compute_clas(img):
|
202 |
+
inputs = clas_processor(images=img, return_tensors="pt")
|
203 |
+
outputs = clas_model(**inputs)
|
204 |
+
logits = outputs.logits
|
205 |
+
predicted_class_idx = logits.argmax(-1).item()
|
206 |
+
return[img, clas_model.config.id2label[predicted_class_idx]]
|
207 |
+
|
208 |
+
m2f_processor = AutoImageProcessor.from_pretrained("facebook/mask2former-swin-base-coco-panoptic")
|
209 |
+
m2f_model = Mask2FormerForUniversalSegmentation.from_pretrained("facebook/mask2former-swin-base-coco-panoptic")
|
210 |
+
|
211 |
+
def seg2sem(seg):
|
212 |
+
color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8) # height, width, 3
|
213 |
+
|
214 |
+
handles = []
|
215 |
+
for label, color in enumerate(palette):
|
216 |
+
color_seg[seg == label, :] = color
|
217 |
+
if (seg == label).count_nonzero() > 0:
|
218 |
+
handles.append(m2f_model.config.id2label[label])
|
219 |
+
handles.append(color)
|
220 |
+
|
221 |
+
color_seg = color_seg.astype(np.uint8)
|
222 |
+
image = Image.fromarray(color_seg)
|
223 |
+
|
224 |
+
return [image,handles]
|
225 |
+
|
226 |
+
def seg2pano(seg, segments_info):
|
227 |
+
color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8) # height, width, 3
|
228 |
+
|
229 |
+
handles = []
|
230 |
+
for label, color in enumerate(palette):
|
231 |
+
color_seg[seg == label, :] = color
|
232 |
+
|
233 |
+
color_seg = color_seg.astype(np.uint8)
|
234 |
+
image = Image.fromarray(color_seg)
|
235 |
+
|
236 |
+
instances_counter = defaultdict(int)
|
237 |
+
handles = []
|
238 |
+
for segment in segments_info:
|
239 |
+
segment_id = segment['id']
|
240 |
+
segment_label_id = segment['label_id']
|
241 |
+
segment_label = m2f_model.config.id2label[segment_label_id]
|
242 |
+
label = f"{segment_label}-{instances_counter[segment_label_id]}"
|
243 |
+
instances_counter[segment_label_id] += 1
|
244 |
+
color = palette[segment_id]
|
245 |
+
handles.append(label)
|
246 |
+
handles.append(color)
|
247 |
+
return [image,handles]
|
248 |
+
|
249 |
+
def compute_m2f_sem_seg(img):
|
250 |
+
inputs = m2f_processor(images=img, return_tensors="pt")
|
251 |
+
|
252 |
+
with torch.no_grad():
|
253 |
+
outputs = m2f_model(**inputs)
|
254 |
+
|
255 |
+
seg = m2f_processor.post_process_semantic_segmentation(
|
256 |
+
outputs, target_sizes=[img.size[::-1]]
|
257 |
+
)[0]
|
258 |
+
|
259 |
+
return seg2sem(seg)
|
260 |
+
|
261 |
+
def compute_m2f_pano_seg(img):
|
262 |
+
inputs = m2f_processor(images=img, return_tensors="pt")
|
263 |
+
|
264 |
+
with torch.no_grad():
|
265 |
+
outputs = m2f_model(**inputs)
|
266 |
+
|
267 |
+
seg = m2f_processor.post_process_panoptic_segmentation(
|
268 |
+
outputs, target_sizes=[img.size[::-1]]
|
269 |
+
)[0]
|
270 |
+
|
271 |
+
return seg2pano(seg["segmentation"], seg["segments_info"])
|
272 |
+
|
273 |
+
labels = ["Dinov2 - Depth", "Dinov2 - Classification", "M2F - Semantic Segmentation", "M2F - Panoptic Segmentation"]
|
274 |
+
|
275 |
+
# main function
|
276 |
+
def detect(img, application):
|
277 |
+
if application == labels[0]:
|
278 |
+
return compute_depth(img)
|
279 |
+
elif application == labels[1]:
|
280 |
+
return compute_clas(img)
|
281 |
+
elif application == labels[2]:
|
282 |
+
return compute_m2f_sem_seg(img)
|
283 |
+
elif application == labels[3]:
|
284 |
+
return compute_m2f_pano_seg(img)
|
285 |
+
return img
|
286 |
+
|
287 |
+
# visual gradio interface
|
288 |
+
iface = gr.Interface(fn=detect, inputs=[gr.Image(type="pil"), gr.Radio(labels, label="Application")], outputs=[gr.Image(type="pil"), gr.Textbox()])
|
289 |
+
iface.launch(debug=True)
|