ritwikraha
commited on
Commit
•
49e0d56
1
Parent(s):
0bfcfed
add: main script added
Browse files- app.py +267 -0
- nerf/keras_metadata.pb +3 -0
- nerf/saved_model.pb +3 -0
- nerf/variables/variables.data-00000-of-00001 +0 -0
- nerf/variables/variables.index +0 -0
app.py
ADDED
@@ -0,0 +1,267 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
# Setting random seed to obtain reproducible results.
|
3 |
+
import tensorflow as tf
|
4 |
+
|
5 |
+
tf.random.set_seed(42)
|
6 |
+
|
7 |
+
import os
|
8 |
+
import glob
|
9 |
+
import imageio
|
10 |
+
from PIL import Image
|
11 |
+
import numpy as np
|
12 |
+
from tqdm import tqdm
|
13 |
+
from tensorflow import keras
|
14 |
+
from tensorflow.keras import layers
|
15 |
+
import matplotlib.pyplot as plt
|
16 |
+
|
17 |
+
# Initialize global variables.
|
18 |
+
AUTO = tf.data.AUTOTUNE
|
19 |
+
BATCH_SIZE = 1
|
20 |
+
NUM_SAMPLES = 32
|
21 |
+
POS_ENCODE_DIMS = 16
|
22 |
+
EPOCHS = 20
|
23 |
+
H = 100
|
24 |
+
W = 100
|
25 |
+
focal = 138.88
|
26 |
+
|
27 |
+
def encode_position(x):
|
28 |
+
"""Encodes the position into its corresponding Fourier feature.
|
29 |
+
|
30 |
+
Args:
|
31 |
+
x: The input coordinate.
|
32 |
+
|
33 |
+
Returns:
|
34 |
+
Fourier features tensors of the position.
|
35 |
+
"""
|
36 |
+
positions = [x]
|
37 |
+
for i in range(POS_ENCODE_DIMS):
|
38 |
+
for fn in [tf.sin, tf.cos]:
|
39 |
+
positions.append(fn(2.0 ** i * x))
|
40 |
+
return tf.concat(positions, axis=-1)
|
41 |
+
|
42 |
+
|
43 |
+
def get_rays(height, width, focal, pose):
|
44 |
+
"""Computes origin point and direction vector of rays.
|
45 |
+
|
46 |
+
Args:
|
47 |
+
height: Height of the image.
|
48 |
+
width: Width of the image.
|
49 |
+
focal: The focal length between the images and the camera.
|
50 |
+
pose: The pose matrix of the camera.
|
51 |
+
|
52 |
+
Returns:
|
53 |
+
Tuple of origin point and direction vector for rays.
|
54 |
+
"""
|
55 |
+
# Build a meshgrid for the rays.
|
56 |
+
i, j = tf.meshgrid(
|
57 |
+
tf.range(width, dtype=tf.float32),
|
58 |
+
tf.range(height, dtype=tf.float32),
|
59 |
+
indexing="xy",
|
60 |
+
)
|
61 |
+
|
62 |
+
# Normalize the x axis coordinates.
|
63 |
+
transformed_i = (i - width * 0.5) / focal
|
64 |
+
|
65 |
+
# Normalize the y axis coordinates.
|
66 |
+
transformed_j = (j - height * 0.5) / focal
|
67 |
+
|
68 |
+
# Create the direction unit vectors.
|
69 |
+
directions = tf.stack([transformed_i, -transformed_j, -tf.ones_like(i)], axis=-1)
|
70 |
+
|
71 |
+
# Get the camera matrix.
|
72 |
+
camera_matrix = pose[:3, :3]
|
73 |
+
height_width_focal = pose[:3, -1]
|
74 |
+
|
75 |
+
# Get origins and directions for the rays.
|
76 |
+
transformed_dirs = directions[..., None, :]
|
77 |
+
camera_dirs = transformed_dirs * camera_matrix
|
78 |
+
ray_directions = tf.reduce_sum(camera_dirs, axis=-1)
|
79 |
+
ray_origins = tf.broadcast_to(height_width_focal, tf.shape(ray_directions))
|
80 |
+
|
81 |
+
# Return the origins and directions.
|
82 |
+
return (ray_origins, ray_directions)
|
83 |
+
|
84 |
+
|
85 |
+
def render_flat_rays(ray_origins, ray_directions, near, far, num_samples, rand=False):
|
86 |
+
"""Renders the rays and flattens it.
|
87 |
+
|
88 |
+
Args:
|
89 |
+
ray_origins: The origin points for rays.
|
90 |
+
ray_directions: The direction unit vectors for the rays.
|
91 |
+
near: The near bound of the volumetric scene.
|
92 |
+
far: The far bound of the volumetric scene.
|
93 |
+
num_samples: Number of sample points in a ray.
|
94 |
+
rand: Choice for randomising the sampling strategy.
|
95 |
+
|
96 |
+
Returns:
|
97 |
+
Tuple of flattened rays and sample points on each rays.
|
98 |
+
"""
|
99 |
+
# Compute 3D query points.
|
100 |
+
# Equation: r(t) = o+td -> Building the "t" here.
|
101 |
+
t_vals = tf.linspace(near, far, num_samples)
|
102 |
+
if rand:
|
103 |
+
# Inject uniform noise into sample space to make the sampling
|
104 |
+
# continuous.
|
105 |
+
shape = list(ray_origins.shape[:-1]) + [num_samples]
|
106 |
+
noise = tf.random.uniform(shape=shape) * (far - near) / num_samples
|
107 |
+
t_vals = t_vals + noise
|
108 |
+
|
109 |
+
# Equation: r(t) = o + td -> Building the "r" here.
|
110 |
+
rays = ray_origins[..., None, :] + (
|
111 |
+
ray_directions[..., None, :] * t_vals[..., None]
|
112 |
+
)
|
113 |
+
rays_flat = tf.reshape(rays, [-1, 3])
|
114 |
+
rays_flat = encode_position(rays_flat)
|
115 |
+
return (rays_flat, t_vals)
|
116 |
+
|
117 |
+
|
118 |
+
def map_fn(pose):
|
119 |
+
"""Maps individual pose to flattened rays and sample points.
|
120 |
+
|
121 |
+
Args:
|
122 |
+
pose: The pose matrix of the camera.
|
123 |
+
|
124 |
+
Returns:
|
125 |
+
Tuple of flattened rays and sample points corresponding to the
|
126 |
+
camera pose.
|
127 |
+
"""
|
128 |
+
(ray_origins, ray_directions) = get_rays(height=H, width=W, focal=focal, pose=pose)
|
129 |
+
(rays_flat, t_vals) = render_flat_rays(
|
130 |
+
ray_origins=ray_origins,
|
131 |
+
ray_directions=ray_directions,
|
132 |
+
near=2.0,
|
133 |
+
far=6.0,
|
134 |
+
num_samples=NUM_SAMPLES,
|
135 |
+
rand=True,
|
136 |
+
)
|
137 |
+
return (rays_flat, t_vals)
|
138 |
+
|
139 |
+
def render_rgb_depth(model, rays_flat, t_vals, rand=True, train=True):
|
140 |
+
"""Generates the RGB image and depth map from model prediction.
|
141 |
+
|
142 |
+
Args:
|
143 |
+
model: The MLP model that is trained to predict the rgb and
|
144 |
+
volume density of the volumetric scene.
|
145 |
+
rays_flat: The flattened rays that serve as the input to
|
146 |
+
the NeRF model.
|
147 |
+
t_vals: The sample points for the rays.
|
148 |
+
rand: Choice to randomise the sampling strategy.
|
149 |
+
train: Whether the model is in the training or testing phase.
|
150 |
+
|
151 |
+
Returns:
|
152 |
+
Tuple of rgb image and depth map.
|
153 |
+
"""
|
154 |
+
# Get the predictions from the nerf model and reshape it.
|
155 |
+
if train:
|
156 |
+
predictions = model(rays_flat)
|
157 |
+
else:
|
158 |
+
predictions = model.predict(rays_flat)
|
159 |
+
predictions = tf.reshape(predictions, shape=(BATCH_SIZE, H, W, NUM_SAMPLES, 4))
|
160 |
+
|
161 |
+
# Slice the predictions into rgb and sigma.
|
162 |
+
rgb = tf.sigmoid(predictions[..., :-1])
|
163 |
+
sigma_a = tf.nn.relu(predictions[..., -1])
|
164 |
+
|
165 |
+
# Get the distance of adjacent intervals.
|
166 |
+
delta = t_vals[..., 1:] - t_vals[..., :-1]
|
167 |
+
# delta shape = (num_samples)
|
168 |
+
if rand:
|
169 |
+
delta = tf.concat(
|
170 |
+
[delta, tf.broadcast_to([1e10], shape=(BATCH_SIZE, H, W, 1))], axis=-1
|
171 |
+
)
|
172 |
+
alpha = 1.0 - tf.exp(-sigma_a * delta)
|
173 |
+
else:
|
174 |
+
delta = tf.concat(
|
175 |
+
[delta, tf.broadcast_to([1e10], shape=(BATCH_SIZE, 1))], axis=-1
|
176 |
+
)
|
177 |
+
alpha = 1.0 - tf.exp(-sigma_a * delta[:, None, None, :])
|
178 |
+
|
179 |
+
# Get transmittance.
|
180 |
+
exp_term = 1.0 - alpha
|
181 |
+
epsilon = 1e-10
|
182 |
+
transmittance = tf.math.cumprod(exp_term + epsilon, axis=-1, exclusive=True)
|
183 |
+
weights = alpha * transmittance
|
184 |
+
rgb = tf.reduce_sum(weights[..., None] * rgb, axis=-2)
|
185 |
+
|
186 |
+
if rand:
|
187 |
+
depth_map = tf.reduce_sum(weights * t_vals, axis=-1)
|
188 |
+
else:
|
189 |
+
depth_map = tf.reduce_sum(weights * t_vals[:, None, None], axis=-1)
|
190 |
+
return (rgb, depth_map)
|
191 |
+
|
192 |
+
nerf_loaded = tf.keras.models.load_model("nerf", compile=False)
|
193 |
+
|
194 |
+
def get_translation_t(t):
|
195 |
+
"""Get the translation matrix for movement in t."""
|
196 |
+
matrix = [
|
197 |
+
[1, 0, 0, 0],
|
198 |
+
[0, 1, 0, 0],
|
199 |
+
[0, 0, 1, t],
|
200 |
+
[0, 0, 0, 1],
|
201 |
+
]
|
202 |
+
return tf.convert_to_tensor(matrix, dtype=tf.float32)
|
203 |
+
|
204 |
+
|
205 |
+
def get_rotation_phi(phi):
|
206 |
+
"""Get the rotation matrix for movement in phi."""
|
207 |
+
matrix = [
|
208 |
+
[1, 0, 0, 0],
|
209 |
+
[0, tf.cos(phi), -tf.sin(phi), 0],
|
210 |
+
[0, tf.sin(phi), tf.cos(phi), 0],
|
211 |
+
[0, 0, 0, 1],
|
212 |
+
]
|
213 |
+
return tf.convert_to_tensor(matrix, dtype=tf.float32)
|
214 |
+
|
215 |
+
|
216 |
+
def get_rotation_theta(theta):
|
217 |
+
"""Get the rotation matrix for movement in theta."""
|
218 |
+
matrix = [
|
219 |
+
[tf.cos(theta), 0, -tf.sin(theta), 0],
|
220 |
+
[0, 1, 0, 0],
|
221 |
+
[tf.sin(theta), 0, tf.cos(theta), 0],
|
222 |
+
[0, 0, 0, 1],
|
223 |
+
]
|
224 |
+
return tf.convert_to_tensor(matrix, dtype=tf.float32)
|
225 |
+
|
226 |
+
|
227 |
+
def pose_spherical(theta, phi, t):
|
228 |
+
"""
|
229 |
+
Get the camera to world matrix for the corresponding theta, phi
|
230 |
+
and t.
|
231 |
+
"""
|
232 |
+
c2w = get_translation_t(t)
|
233 |
+
c2w = get_rotation_phi(phi / 180.0 * np.pi) @ c2w
|
234 |
+
c2w = get_rotation_theta(theta / 180.0 * np.pi) @ c2w
|
235 |
+
c2w = np.array([[-1, 0, 0, 0], [0, 0, 1, 0], [0, 1, 0, 0], [0, 0, 0, 1]]) @ c2w
|
236 |
+
return c2w
|
237 |
+
|
238 |
+
|
239 |
+
def show_rendered_image(r,theta,phi):
|
240 |
+
# Get the camera to world matrix.
|
241 |
+
c2w = pose_spherical(theta, phi, r)
|
242 |
+
|
243 |
+
ray_oris, ray_dirs = get_rays(H, W, focal, c2w)
|
244 |
+
rays_flat, t_vals = render_flat_rays(
|
245 |
+
ray_oris, ray_dirs, near=2.0, far=6.0, num_samples=NUM_SAMPLES, rand=False
|
246 |
+
)
|
247 |
+
|
248 |
+
rgb, depth = render_rgb_depth(
|
249 |
+
nerf_loaded, rays_flat[None, ...], t_vals[None, ...], rand=False, train=False
|
250 |
+
)
|
251 |
+
return(rgb[0], depth[0])
|
252 |
+
|
253 |
+
# app.py text matter starts here
|
254 |
+
st.title('NeRF:Neural Radiance Fields')
|
255 |
+
st.subfield('')
|
256 |
+
# set the values of r theta phi
|
257 |
+
r = -30.0
|
258 |
+
theta = st.slider('Enter a value for theta', 0.0, 360.0, 1)
|
259 |
+
phi = st.slider('Enter a value for phi', 0.0, 360.0, 1)
|
260 |
+
|
261 |
+
color,depth = show_rendered_image(r,theta,phi)
|
262 |
+
|
263 |
+
st.image(color, caption = "Color")
|
264 |
+
st.image(depth, caption = "Depth")
|
265 |
+
|
266 |
+
|
267 |
+
|
nerf/keras_metadata.pb
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e8da49eeec070f24b87d869c2005bdec6fdbd1a1bc1fb6d44c73eb8f89321c6c
|
3 |
+
size 21754
|
nerf/saved_model.pb
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9fe43d8f799d56fc7ecdc964172f56541f7cbdaed8644559ed9c7bac553e826e
|
3 |
+
size 272106
|
nerf/variables/variables.data-00000-of-00001
ADDED
Binary file (174 kB). View file
|
|
nerf/variables/variables.index
ADDED
Binary file (1.24 kB). View file
|
|