nerf_helpers.py

import tensorflow as tf
import numpy as np


# Misc utils

img2mse = lambda x, y: tf.reduce_mean(tf.square(x - y))
img2mae = lambda x, y: tf.reduce_mean(tf.abs(x - y))
mse2psnr = lambda x: -10.*tf.math.log(x)/tf.math.log(10.)
to8b = lambda x: (255*np.clip(x, 0, 1)).astype(np.uint8)
to_depth16 = lambda x: (1000 * x).astype(np.uint16)

# Positional encoding

class Embedder:

    def __init__(self, **kwargs):
        self.kwargs = kwargs
        self.create_embedding_fn()

    def create_embedding_fn(self):

        embed_fns = []
        d = self.kwargs['input_dims']
        out_dim = 0
        if self.kwargs['include_input']:
            embed_fns.append(lambda x: x)
            out_dim += d

        max_freq = self.kwargs['max_freq_log2']
        N_freqs = self.kwargs['num_freqs']

        if self.kwargs['log_sampling']:
            freq_bands = 2.**tf.linspace(0., max_freq, N_freqs)
        else:
            freq_bands = tf.linspace(2.**0., 2.**max_freq, N_freqs)

        if self.kwargs['gaussian']:
            B = tf.random.normal([3, 256], 0.0, 10.0, seed=0)

        for freq in freq_bands:
            for p_fn in self.kwargs['periodic_fns']:
                if self.kwargs['gaussian']:
                    embed_fns.append(lambda x, p_fn=p_fn, freq=freq, B=B: p_fn(tf.matmul(x, B) * freq))
                else:
                    embed_fns.append(lambda x, p_fn=p_fn, freq=freq: p_fn(x * freq))
                out_dim += d

        self.embed_fns = embed_fns
        self.out_dim = out_dim

        if self.kwargs['gaussian']:
            self.out_dim = B.shape[1] * 2

    def embed(self, inputs):
        return tf.concat([fn(inputs) for fn in self.embed_fns], -1)


def get_embedder_obj(multires):
    embed_kwargs = {
                'include_input': True,
                'input_dims': 3,
                'max_freq_log2': multires-1,
                'num_freqs': multires,
                'log_sampling': True,
                'periodic_fns': [tf.math.sin, tf.math.cos],
                'gaussian': False,
    }

    return Embedder(**embed_kwargs)


def get_embedder(multires, i=0):

    if i == -1:
        return tf.identity, 3

    embedder_obj = get_embedder_obj(multires)
    embed = lambda x, eo=embedder_obj: eo.embed(x)
    return embed, embedder_obj.out_dim


# Model architecture

def init_nerf_model(D=8, W=256, input_ch=3, input_ch_views=3, output_ch=4, skips=[4], use_viewdirs=False):

    relu = tf.keras.layers.ReLU()
    dense = lambda W, act=relu: tf.keras.layers.Dense(W, activation=act)

    print('MODEL', input_ch, input_ch_views, type(input_ch), type(input_ch_views), use_viewdirs)
    input_ch = int(input_ch)
    input_ch_views = int(input_ch_views)

    inputs = tf.keras.Input(shape=(input_ch + input_ch_views))
    inputs_pts, inputs_views = tf.split(inputs, [input_ch, input_ch_views], -1)
    inputs_pts.set_shape([None, input_ch])
    inputs_views.set_shape([None, input_ch_views])

    print(inputs.shape, inputs_pts.shape, inputs_views.shape)
    outputs = inputs_pts
    for i in range(D):
        outputs = dense(W)(outputs)
        if i in skips:
            outputs = tf.concat([inputs_pts, outputs], -1)

    if use_viewdirs:
        alpha_out = dense(1, act=None)(outputs)
        bottleneck = dense(256, act=None)(outputs)
        inputs_viewdirs = tf.concat([bottleneck, inputs_views], -1)  # concat viewdirs
        outputs = inputs_viewdirs
        for i in range(1):
            outputs = dense(W//2)(outputs)
        outputs = dense(3, act=None)(outputs)
        outputs = tf.concat([outputs, alpha_out], -1)
    else:
        outputs = dense(output_ch, act=None)(outputs)

    model = tf.keras.Model(inputs=inputs, outputs=outputs)
    return model


# Ray helpers

def get_rays(H, W, focal, c2w):
    """Get ray origins, directions from a pinhole camera."""
    i, j = tf.meshgrid(tf.range(W, dtype=tf.float32),
                       tf.range(H, dtype=tf.float32), indexing='xy')
    dirs = tf.stack([(i + 0.5 - W*.5)/focal, -(j + 0.5 - H*.5)/focal, -tf.ones_like(i)], -1)
    rays_d = tf.reduce_sum(dirs[..., np.newaxis, :] * c2w[:3, :3], -1)
    rays_o = tf.broadcast_to(c2w[:3, -1], tf.shape(rays_d))
    return rays_o, rays_d


def get_rays_np(H, W, focal, c2w):
    """Get ray origins, directions from a pinhole camera."""
    i, j = np.meshgrid(np.arange(W, dtype=np.float32),
                       np.arange(H, dtype=np.float32), indexing='xy')
    dirs = np.stack([(i + 0.5 - W*.5)/focal, -(j + 0.5 - H*.5)/focal, -np.ones_like(i)], -1)
    rays_d = np.sum(dirs[..., np.newaxis, :] * c2w[:3, :3], -1)
    rays_o = np.broadcast_to(c2w[:3, -1], np.shape(rays_d))
    return rays_o, rays_d


def get_camera_rays_np(H, W, focal):
    """Get ray origins, directions from a pinhole camera."""
    i, j = np.meshgrid(np.arange(W, dtype=np.float32),
                       np.arange(H, dtype=np.float32), indexing='xy')
    dirs = np.stack([(i + 0.5 - W*.5)/focal, -(j + 0.5 - H*.5)/focal, -np.ones_like(i)], -1)
    rays_d = dirs
    return rays_d


def get_rays_np_random(H, W, focal, c2w):
    """Get ray origins, directions from a pinhole camera."""
    i, j = np.meshgrid(np.arange(W, dtype=np.float32),
                       np.arange(H, dtype=np.float32), indexing='xy')

    i_rand = np.random.rand(*i.shape)
    j_rand = np.random.rand(*j.shape)

    dirs = np.stack([(i + i_rand - W*.5)/focal, -(j + j_rand - H*.5)/focal, -np.ones_like(i)], -1)
    rays_d = np.sum(dirs[..., np.newaxis, :] * c2w[:3, :3], -1)
    rays_o = np.broadcast_to(c2w[:3, -1], np.shape(rays_d))
    return rays_o, rays_d


def ndc_rays(H, W, focal, near, rays_o, rays_d):
    """Normalized device coordinate rays.
    Space such that the canvas is a cube with sides [-1, 1] in each axis.
    Args:
      H: int. Height in pixels.
      W: int. Width in pixels.
      focal: float. Focal length of pinhole camera.
      near: float or array of shape[batch_size]. Near depth bound for the scene.
      rays_o: array of shape [batch_size, 3]. Camera origin.
      rays_d: array of shape [batch_size, 3]. Ray direction.
    Returns:
      rays_o: array of shape [batch_size, 3]. Camera origin in NDC.
      rays_d: array of shape [batch_size, 3]. Ray direction in NDC.
    """
    # Shift ray origins to near plane
    t = -(near + rays_o[..., 2]) / rays_d[..., 2]
    rays_o = rays_o + t[..., None] * rays_d

    # Projection
    o0 = -1./(W/(2.*focal)) * rays_o[..., 0] / rays_o[..., 2]
    o1 = -1./(H/(2.*focal)) * rays_o[..., 1] / rays_o[..., 2]
    o2 = 1. + 2. * near / rays_o[..., 2]

    d0 = -1./(W/(2.*focal)) * \
        (rays_d[..., 0]/rays_d[..., 2] - rays_o[..., 0]/rays_o[..., 2])
    d1 = -1./(H/(2.*focal)) * \
        (rays_d[..., 1]/rays_d[..., 2] - rays_o[..., 1]/rays_o[..., 2])
    d2 = -2. * near / rays_o[..., 2]

    rays_o = tf.stack([o0, o1, o2], -1)
    rays_d = tf.stack([d0, d1, d2], -1)

    return rays_o, rays_d


# Hierarchical sampling helper

def sample_pdf(bins, weights, N_samples, det=False):

    # Get pdf
    weights += 1e-5  # prevent nans
    pdf = weights / tf.reduce_sum(weights, -1, keepdims=True)
    cdf = tf.cumsum(pdf, -1)
    cdf = tf.concat([tf.zeros_like(cdf[..., :1]), cdf], -1)

    # Take uniform samples
    if det:
        u = tf.linspace(0., 1., N_samples)
        u = tf.broadcast_to(u, list(cdf.shape[:-1]) + [N_samples])
    else:
        u = tf.random.uniform(list(cdf.shape[:-1]) + [N_samples])

    # Invert CDF
    inds = tf.searchsorted(cdf, u, side='right')
    below = tf.maximum(0, inds-1)
    above = tf.minimum(cdf.shape[-1]-1, inds)
    inds_g = tf.stack([below, above], -1)
    cdf_g = tf.gather(cdf, inds_g, axis=-1, batch_dims=len(inds_g.shape)-2)
    bins_g = tf.gather(bins, inds_g, axis=-1, batch_dims=len(inds_g.shape)-2)

    denom = (cdf_g[..., 1] - cdf_g[..., 0])
    denom = tf.where(denom < 1e-5, tf.ones_like(denom), denom)
    t = (u-cdf_g[..., 0]) / denom
    samples = bins_g[..., 0] + t * (bins_g[..., 1] - bins_g[..., 0])

    return samples