Spaces:
Running on Zero
Running on Zero
File size: 9,430 Bytes
7405e85 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 | import torch
import torch.nn.functional as F
from einops import rearrange, repeat
from torch import nn
from ....geometry.projection import (
homogenize_points,
project,
sample_image_grid,
transform_cam2world,
transform_world2cam,
unproject,
)
class DepthGuidedMVRefiner(nn.Module):
def __init__(
self,
feature_channels: int,
hidden_channels: int | None = None,
concat_ref_depth: bool = False,
) -> None:
super().__init__()
hidden_channels = feature_channels if hidden_channels is None else hidden_channels
depth_channels = max(hidden_channels // 4, 16)
self.concat_ref_depth = concat_ref_depth
self.depth_encoder = nn.Sequential(
nn.Conv2d(
1,
depth_channels,
1,
1,
0,
),
nn.GELU(),
nn.Conv2d(
depth_channels,
depth_channels,
1,
1,
0,
),
nn.GELU(),
)
self.feature_encoder = nn.Sequential(
nn.Conv2d(
feature_channels * 2
+ depth_channels * (2 if self.concat_ref_depth else 1),
hidden_channels,
3,
1,
1,
padding_mode="replicate",
),
nn.GELU(),
nn.Conv2d(
hidden_channels,
hidden_channels,
3,
1,
1,
padding_mode="replicate",
),
nn.GELU(),
)
self.post_fuse = nn.Sequential(
nn.Conv2d(
feature_channels + hidden_channels + depth_channels * int(self.concat_ref_depth),
hidden_channels,
3,
1,
1,
padding_mode="replicate",
),
nn.GELU(),
nn.Conv2d(
hidden_channels,
feature_channels,
3,
1,
1,
padding_mode="replicate",
),
)
nn.init.zeros_(self.post_fuse[-1].weight)
nn.init.zeros_(self.post_fuse[-1].bias)
@staticmethod
def _normalized_coords_to_grid(
xy: torch.Tensor,
h: int,
w: int,
) -> tuple[torch.Tensor, torch.Tensor]:
x_pixel = xy[..., 0] * w - 0.5
y_pixel = xy[..., 1] * h - 0.5
if w > 1:
x_grid = 2 * x_pixel / (w - 1) - 1
else:
x_grid = torch.zeros_like(x_pixel)
if h > 1:
y_grid = 2 * y_pixel / (h - 1) - 1
else:
y_grid = torch.zeros_like(y_pixel)
grid = torch.stack((x_grid, y_grid), dim=-1)
in_bounds = (
(x_pixel >= 0)
& (x_pixel <= (w - 1))
& (y_pixel >= 0)
& (y_pixel <= (h - 1))
)
return grid, in_bounds
def forward(
self,
features: torch.Tensor,
depth: torch.Tensor,
intrinsics: torch.Tensor,
extrinsics: torch.Tensor,
source_view_indices: torch.Tensor | None = None,
) -> torch.Tensor:
b, v, c, h, w = features.shape
if v <= 1:
return features
depth = depth.detach()
intrinsics = intrinsics.detach()
extrinsics = extrinsics.detach()
feature_dtype = features.dtype
xy_ref, _ = sample_image_grid((h, w), device=features.device)
xy_ref = repeat(xy_ref, "h w xy -> b h w xy", b=b)
refined_features = []
geom_eps = torch.finfo(torch.float32).eps
count_eps = torch.finfo(features.dtype).eps
batch_indices = torch.arange(b, device=features.device)
if source_view_indices is not None:
if source_view_indices.shape[:2] != (b, v):
raise ValueError(
"source_view_indices must have shape [B, V, M], got "
f"{tuple(source_view_indices.shape)} for features "
f"{tuple(features.shape)}"
)
source_view_indices = source_view_indices.to(
device=features.device,
dtype=torch.long,
)
for ref_idx in range(v):
ref_features = features[:, ref_idx]
ref_depth = depth[:, ref_idx].float()
ref_intrinsics = intrinsics[:, ref_idx].float()
ref_extrinsics = extrinsics[:, ref_idx].float()
ref_depth_repr = None
if self.concat_ref_depth:
ref_depth_input = ref_depth.unsqueeze(1).clamp_min(geom_eps).reciprocal()
ref_depth_repr = self.depth_encoder(ref_depth_input.to(feature_dtype))
with torch.no_grad(), torch.autocast(
device_type=features.device.type, enabled=False
):
ref_points_cam = unproject(
xy_ref,
ref_depth,
ref_intrinsics[:, None, None],
)
ref_points_world = transform_cam2world(
homogenize_points(ref_points_cam),
ref_extrinsics[:, None, None],
)[..., :-1]
fused_sum = None
valid_count = None
if source_view_indices is None:
selected_src_indices = [
src_idx for src_idx in range(v) if src_idx != ref_idx
]
else:
selected_src_indices = source_view_indices[:, ref_idx].unbind(dim=-1)
for src_idx in selected_src_indices:
if isinstance(src_idx, int):
src_features = features[:, src_idx]
src_extrinsics = extrinsics[:, src_idx].float()[:, None, None]
src_intrinsics = intrinsics[:, src_idx].float()[:, None, None]
else:
src_features = features[batch_indices, src_idx]
src_extrinsics = (
extrinsics[batch_indices, src_idx].float()[:, None, None]
)
src_intrinsics = (
intrinsics[batch_indices, src_idx].float()[:, None, None]
)
with torch.no_grad(), torch.autocast(
device_type=features.device.type, enabled=False
):
src_xy, src_in_front = project(
ref_points_world,
src_extrinsics,
src_intrinsics,
)
src_grid, src_in_bounds = self._normalized_coords_to_grid(
src_xy, h, w
)
valid_mask_fp32 = (src_in_front & src_in_bounds).unsqueeze(1).float()
with torch.autocast(device_type=features.device.type, enabled=False):
warped_features = F.grid_sample(
src_features.float(),
src_grid,
mode="bilinear",
padding_mode="zeros",
align_corners=True,
)
valid_mask = valid_mask_fp32.to(feature_dtype)
warped_features = warped_features.to(feature_dtype)
warped_features = warped_features * valid_mask
with torch.no_grad(), torch.autocast(
device_type=features.device.type, enabled=False
):
src_points_cam = transform_world2cam(
homogenize_points(ref_points_world),
src_extrinsics,
)[..., :-1]
warped_depth = rearrange(
src_points_cam[..., -1], "b h w -> b 1 h w"
)
warped_depth_input = torch.where(
valid_mask_fp32 > 0,
warped_depth.clamp_min(geom_eps).reciprocal(),
torch.zeros_like(warped_depth),
)
depth_repr = self.depth_encoder(warped_depth_input.to(feature_dtype))
feature_inputs = [ref_features, warped_features]
if ref_depth_repr is not None:
feature_inputs.append(ref_depth_repr)
feature_inputs.append(depth_repr)
value = self.feature_encoder(torch.cat(feature_inputs, dim=1)) * valid_mask
if fused_sum is None:
fused_sum = value
valid_count = valid_mask
else:
fused_sum = fused_sum + value
valid_count = valid_count + valid_mask
if fused_sum is None or valid_count is None:
refined_features.append(ref_features)
continue
fused = fused_sum / valid_count.clamp_min(count_eps)
post_fuse_inputs = [ref_features, fused]
if ref_depth_repr is not None:
post_fuse_inputs.append(ref_depth_repr)
residual = self.post_fuse(torch.cat(post_fuse_inputs, dim=1))
refined_features.append(ref_features + residual)
return torch.stack(refined_features, dim=1)
|