File size: 7,137 Bytes
cd769b1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import itertools
import torch
import torch.nn as nn
from .renderer import ImportanceRenderer
from .ray_sampler_part import RaySampler
class OSGDecoder(nn.Module):
"""
Triplane decoder that gives RGB and sigma values from sampled features.
Using ReLU here instead of Softplus in the original implementation.
Reference:
EG3D: https://github.com/NVlabs/eg3d/blob/main/eg3d/training/triplane.py#L112
"""
def __init__(self, n_features: int,
hidden_dim: int = 64, num_layers: int = 4, activation: nn.Module = nn.ReLU):
super().__init__()
self.net = nn.Sequential(
nn.Linear(3 * n_features, hidden_dim),
activation(),
*itertools.chain(*[[
nn.Linear(hidden_dim, hidden_dim),
activation(),
] for _ in range(num_layers - 2)]),
nn.Linear(hidden_dim, 1 + 3),
)
# init all bias to zero
for m in self.modules():
if isinstance(m, nn.Linear):
nn.init.zeros_(m.bias)
def forward(self, sampled_features, ray_directions):
# Aggregate features by mean
# sampled_features = sampled_features.mean(1)
# Aggregate features by concatenation
_N, n_planes, _M, _C = sampled_features.shape
sampled_features = sampled_features.permute(0, 2, 1, 3).reshape(_N, _M, n_planes*_C)
x = sampled_features
N, M, C = x.shape
x = x.contiguous().view(N*M, C)
x = self.net(x)
x = x.view(N, M, -1)
rgb = torch.sigmoid(x[..., 1:])*(1 + 2*0.001) - 0.001 # Uses sigmoid clamping from MipNeRF
sigma = x[..., 0:1]
return {'rgb': rgb, 'sigma': sigma}
class TriplaneSynthesizer(nn.Module):
"""
Synthesizer that renders a triplane volume with planes and a camera.
Reference:
EG3D: https://github.com/NVlabs/eg3d/blob/main/eg3d/training/triplane.py#L19
"""
DEFAULT_RENDERING_KWARGS = {
'ray_start': 'auto',
'ray_end': 'auto',
'box_warp': 2.,
'white_back': True,
'disparity_space_sampling': False,
'clamp_mode': 'softplus',
'sampler_bbox_min': -1.,
'sampler_bbox_max': 1.,
}
def __init__(self, triplane_dim: int, samples_per_ray: int):
super().__init__()
# attributes
self.triplane_dim = triplane_dim
self.rendering_kwargs = {
**self.DEFAULT_RENDERING_KWARGS,
'depth_resolution': samples_per_ray // 2,
'depth_resolution_importance': samples_per_ray // 2,
}
# renderings
self.renderer = ImportanceRenderer()
self.ray_sampler = RaySampler()
# modules
self.decoder = OSGDecoder(n_features=triplane_dim)
def forward(self, planes, cameras, render_size: int, crop_size: int, start_x: int, start_y:int):
# planes: (N, 3, D', H', W')
# cameras: (N, M, D_cam)
# render_size: int
assert planes.shape[0] == cameras.shape[0], "Batch size mismatch for planes and cameras"
N, M = cameras.shape[:2]
cam2world_matrix = cameras[..., :16].view(N, M, 4, 4)
intrinsics = cameras[..., 16:25].view(N, M, 3, 3)
# Create a batch of rays for volume rendering
ray_origins, ray_directions = self.ray_sampler(
cam2world_matrix=cam2world_matrix.reshape(-1, 4, 4),
intrinsics=intrinsics.reshape(-1, 3, 3),
render_size=render_size,
crop_size = crop_size,
start_x = start_x,
start_y = start_y
)
assert N*M == ray_origins.shape[0], "Batch size mismatch for ray_origins"
assert ray_origins.dim() == 3, "ray_origins should be 3-dimensional"
# Perform volume rendering
rgb_samples, depth_samples, weights_samples = self.renderer(
planes.repeat_interleave(M, dim=0), self.decoder, ray_origins, ray_directions, self.rendering_kwargs,
)
# Reshape into 'raw' neural-rendered image
Himg = Wimg = crop_size
rgb_images = rgb_samples.permute(0, 2, 1).reshape(N, M, rgb_samples.shape[-1], Himg, Wimg).contiguous()
depth_images = depth_samples.permute(0, 2, 1).reshape(N, M, 1, Himg, Wimg)
weight_images = weights_samples.permute(0, 2, 1).reshape(N, M, 1, Himg, Wimg)
return {
'images_rgb': rgb_images,
'images_depth': depth_images,
'images_weight': weight_images,
}
def forward_grid(self, planes, grid_size: int, aabb: torch.Tensor = None):
# planes: (N, 3, D', H', W')
# grid_size: int
# aabb: (N, 2, 3)
if aabb is None:
aabb = torch.tensor([
[self.rendering_kwargs['sampler_bbox_min']] * 3,
[self.rendering_kwargs['sampler_bbox_max']] * 3,
], device=planes.device, dtype=planes.dtype).unsqueeze(0).repeat(planes.shape[0], 1, 1)
assert planes.shape[0] == aabb.shape[0], "Batch size mismatch for planes and aabb"
N = planes.shape[0]
# create grid points for triplane query
grid_points = []
for i in range(N):
grid_points.append(torch.stack(torch.meshgrid(
torch.linspace(aabb[i, 0, 0], aabb[i, 1, 0], grid_size, device=planes.device),
torch.linspace(aabb[i, 0, 1], aabb[i, 1, 1], grid_size, device=planes.device),
torch.linspace(aabb[i, 0, 2], aabb[i, 1, 2], grid_size, device=planes.device),
indexing='ij',
), dim=-1).reshape(-1, 3))
cube_grid = torch.stack(grid_points, dim=0).to(planes.device)
features = self.forward_points(planes, cube_grid)
# reshape into grid
features = {
k: v.reshape(N, grid_size, grid_size, grid_size, -1)
for k, v in features.items()
}
return features
def forward_points(self, planes, points: torch.Tensor, chunk_size: int = 2**20):
# planes: (N, 3, D', H', W')
# points: (N, P, 3)
N, P = points.shape[:2]
# query triplane in chunks
outs = []
for i in range(0, points.shape[1], chunk_size):
chunk_points = points[:, i:i+chunk_size]
# query triplane
chunk_out = self.renderer.run_model_activated(
planes=planes,
decoder=self.decoder,
sample_coordinates=chunk_points,
sample_directions=torch.zeros_like(chunk_points),
options=self.rendering_kwargs,
)
outs.append(chunk_out)
# concatenate the outputs
point_features = {
k: torch.cat([out[k] for out in outs], dim=1)
for k in outs[0].keys()
}
sig = point_features['sigma']
print(sig.mean(), sig.max(), sig.min())
return point_features |