rendering.py

import streamlit as st
import tensorflow as tf
import numpy as np

from config import *

def encode_position(x):
    """Encodes the position into its corresponding Fourier feature.
    Args:
        x: The input coordinate.
    Returns:
        Fourier features tensors of the position.
    """
    positions = [x]
    for i in range(POS_ENCODE_DIMS):
        for fn in [tf.sin, tf.cos]:
            positions.append(fn(2.0 ** i * x))
    return tf.concat(positions, axis=-1)


def get_rays(height, width, focal, pose):
    """Computes origin point and direction vector of rays.
    Args:
        height: Height of the image.
        width: Width of the image.
        focal: The focal length between the images and the camera.
        pose: The pose matrix of the camera.
    Returns:
        Tuple of origin point and direction vector for rays.
    """
    # Build a meshgrid for the rays.
    i, j = tf.meshgrid(
        tf.range(width, dtype=tf.float32),
        tf.range(height, dtype=tf.float32),
        indexing="xy",
    )

    # Normalize the x axis coordinates.
    transformed_i = (i - width * 0.5) / focal

    # Normalize the y axis coordinates.
    transformed_j = (j - height * 0.5) / focal

    # Create the direction unit vectors.
    directions = tf.stack([transformed_i, -transformed_j, -tf.ones_like(i)], axis=-1)

    # Get the camera matrix.
    camera_matrix = pose[:3, :3]
    height_width_focal = pose[:3, -1]

    # Get origins and directions for the rays.
    transformed_dirs = directions[..., None, :]
    camera_dirs = transformed_dirs * camera_matrix
    ray_directions = tf.reduce_sum(camera_dirs, axis=-1)
    ray_origins = tf.broadcast_to(height_width_focal, tf.shape(ray_directions))

    # Return the origins and directions.
    return (ray_origins, ray_directions)


def render_flat_rays(ray_origins, ray_directions, near, far, num_samples, rand=False):
    """Renders the rays and flattens it.
    Args:
        ray_origins: The origin points for rays.
        ray_directions: The direction unit vectors for the rays.
        near: The near bound of the volumetric scene.
        far: The far bound of the volumetric scene.
        num_samples: Number of sample points in a ray.
        rand: Choice for randomising the sampling strategy.
    Returns:
       Tuple of flattened rays and sample points on each rays.
    """
    # Compute 3D query points.
    # Equation: r(t) = o+td -> Building the "t" here.
    t_vals = tf.linspace(near, far, num_samples)
    if rand:
        # Inject uniform noise into sample space to make the sampling
        # continuous.
        shape = list(ray_origins.shape[:-1]) + [num_samples]
        noise = tf.random.uniform(shape=shape) * (far - near) / num_samples
        t_vals = t_vals + noise

    # Equation: r(t) = o + td -> Building the "r" here.
    rays = ray_origins[..., None, :] + (
        ray_directions[..., None, :] * t_vals[..., None]
    )
    rays_flat = tf.reshape(rays, [-1, 3])
    rays_flat = encode_position(rays_flat)
    return (rays_flat, t_vals)


def map_fn(pose):
    """Maps individual pose to flattened rays and sample points.
    Args:
        pose: The pose matrix of the camera.
    Returns:
        Tuple of flattened rays and sample points corresponding to the
        camera pose.
    """
    (ray_origins, ray_directions) = get_rays(height=H, width=W, focal=focal, pose=pose)
    (rays_flat, t_vals) = render_flat_rays(
        ray_origins=ray_origins,
        ray_directions=ray_directions,
        near=2.0,
        far=6.0,
        num_samples=NUM_SAMPLES,
        rand=True,
    )
    return (rays_flat, t_vals)


def render_rgb_depth(model, rays_flat, t_vals, rand=True, train=True):
    """Generates the RGB image and depth map from model prediction.
    Args:
        model: The MLP model that is trained to predict the rgb and
            volume density of the volumetric scene.
        rays_flat: The flattened rays that serve as the input to
            the NeRF model.
        t_vals: The sample points for the rays.
        rand: Choice to randomise the sampling strategy.
        train: Whether the model is in the training or testing phase.
    Returns:
        Tuple of rgb image and depth map.
    """
    # Get the predictions from the nerf model and reshape it.
    if train:
        predictions = model(rays_flat)
    else:
        predictions = model.predict(rays_flat)
    predictions = tf.reshape(predictions, shape=(BATCH_SIZE, H, W, NUM_SAMPLES, 4))

    # Slice the predictions into rgb and sigma.
    rgb = tf.sigmoid(predictions[..., :-1])
    sigma_a = tf.nn.relu(predictions[..., -1])

    # Get the distance of adjacent intervals.
    delta = t_vals[..., 1:] - t_vals[..., :-1]
    # delta shape = (num_samples)
    if rand:
        delta = tf.concat(
            [delta, tf.broadcast_to([1e10], shape=(BATCH_SIZE, H, W, 1))], axis=-1
        )
        alpha = 1.0 - tf.exp(-sigma_a * delta)
    else:
        delta = tf.concat(
            [delta, tf.broadcast_to([1e10], shape=(BATCH_SIZE, 1))], axis=-1
        )
        alpha = 1.0 - tf.exp(-sigma_a * delta[:, None, None, :])

    # Get transmittance.
    exp_term = 1.0 - alpha
    epsilon = 1e-10
    transmittance = tf.math.cumprod(exp_term + epsilon, axis=-1, exclusive=True)
    weights = alpha * transmittance
    rgb = tf.reduce_sum(weights[..., None] * rgb, axis=-2)

    if rand:
        depth_map = tf.reduce_sum(weights * t_vals, axis=-1)
    else:
        depth_map = tf.reduce_sum(weights * t_vals[:, None, None], axis=-1)
    return (rgb, depth_map)