Spaces:
Paused
Paused
File size: 15,739 Bytes
08b23ce |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 |
# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from third_partys.Video_Depth_Anything.video_depth_anything.video_depth import VideoDepthAnything
import torch
import torch.nn as nn
import numpy as np
import igl
import cv2
import time
import torch.nn.functional as F
from utils.quat_utils import quat_inverse, quat_log, quat_multiply, normalize_quaternion
from pytorch3d.structures import join_meshes_as_scene, join_meshes_as_batch
import os
from pathlib import Path
class DepthModule:
def __init__(self, encoder='vitl', device='cuda', input_size=518, fp32=False):
"""
Initialize the depth loss module with Video Depth Anything
Args:
encoder: 'vitl' or 'vits'
device: device to run the model on
input_size: input size for the model
fp32: whether to use float32 for inference
"""
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
self.input_size = input_size
self.fp32 = fp32
# Initialize model configuration
model_configs = {
'vits': {'encoder': 'vits', 'features': 64, 'out_channels': [48, 96, 192, 384]},
'vitl': {'encoder': 'vitl', 'features': 256, 'out_channels': [256, 512, 1024, 1024]},
}
# Load Video Depth Anything model
self.video_depth_model = VideoDepthAnything(**model_configs[encoder])
self.video_depth_model.load_state_dict(
torch.load(f'./third_partys/Video_Depth_Anything/ckpt/video_depth_anything_{encoder}.pth', map_location='cpu'),
strict=True
)
self.video_depth_model = self.video_depth_model.to(self.device).eval()
for param in self.video_depth_model.parameters():
param.requires_grad = False
def get_depth_maps(self, frames, target_fps=30):
"""
Get depth maps for video frames
"""
depths, _ = self.video_depth_model.infer_video_depth(
frames,
target_fps,
input_size=self.input_size,
device=self.device,
fp32=self.fp32
)
return depths
def save_depth_as_images(depth_np, output_dir='./depth_images'):
"""save depth images"""
os.makedirs(output_dir, exist_ok=True)
for i, depth_map in enumerate(depth_np):
depth_map = depth_map.detach().cpu().numpy()
valid_mask = (depth_map > 0)
if not valid_mask.any():
continue
valid_min = depth_map[valid_mask].min()
valid_max = depth_map[valid_mask].max()
normalized = np.zeros_like(depth_map)
normalized[valid_mask] = 255.0 * (depth_map[valid_mask] - valid_min) / (valid_max - valid_min)
depth_img = normalized.astype(np.uint8)
cv2.imwrite(os.path.join(output_dir, f'depth_{i:04d}.png'), depth_img)
print(f"Save {len(depth_np)} depth images to {output_dir}")
def compute_visibility_mask_igl(ray_origins, ray_dirs, distances, verts, faces, distance_tolerance=1e-6, for_vertices=False):
"""
Compute visibility mask using IGL ray-mesh intersection.
"""
num_rays = ray_origins.shape[0]
visibility_mask = np.ones(num_rays, dtype=bool)
for i in range(num_rays):
ray_origin = ray_origins[i].reshape(1, 3)
ray_dir = ray_dirs[i].reshape(1, 3)
intersections = igl.ray_mesh_intersect(ray_origin, ray_dir, verts, faces)
if intersections:
# Count intersections that occur before the target point
count = sum(1 for h in intersections if h[4] < distances[i] - distance_tolerance)
# count=0 → ray completely missed the mesh; count=1 → ray stops exactly at the face containing the joint
# count>1 → ray was blocked by other faces along the way
if for_vertices:
if count != 1:
visibility_mask[i] = False
else: # for joints
if count > 2:
visibility_mask[i] = False
return visibility_mask
def compute_reprojection_loss(renderer, vis_mask, predicted_joints, tracked_joints_2d, image_size):
"""
Compute reprojection loss between predicted 3D points and tracked 2D points.
"""
if predicted_joints.dim() != 3:
raise ValueError(f"predicted_joints must be 3D tensor, got shape {predicted_joints.shape}")
B, J, _ = predicted_joints.shape
device = predicted_joints.device
# Project 3D joints to 2D screen coordinates
projected = renderer.camera.transform_points_screen(
predicted_joints,
image_size=[image_size, image_size]
)
projected_2d = projected[..., :2] # (B, J, 2)
# Convert and validate tracked joints
if not isinstance(tracked_joints_2d, torch.Tensor):
tracked_joints_2d = torch.from_numpy(tracked_joints_2d).float()
tracked_joints_2d = tracked_joints_2d.to(device)
if tracked_joints_2d.dim() == 2:
tracked_joints_2d = tracked_joints_2d.unsqueeze(0).expand(B, -1, -1)
vis_mask = vis_mask.to(device).float()
num_visible = vis_mask.sum()
if num_visible == 0:
# No visible joints - return zero loss
return torch.tensor(0.0, device=device, requires_grad=True)
squared_diff = (projected_2d - tracked_joints_2d).pow(2).sum(dim=-1) # (B, J)
vis_mask_expanded = vis_mask.unsqueeze(0) # (1, J)
masked_loss = squared_diff * vis_mask_expanded # (B, J)
per_frame_loss = masked_loss.sum(dim=1) / num_visible # (B,)
final_loss = per_frame_loss.mean() # scalar
return final_loss
def geodesic_loss(q1, q2, eps=1e-6):
"""
Compute geodesic distance loss between batches of quaternions for rot smooth loss.
"""
q1_norm = normalize_quaternion(q1, eps=eps)
q2_norm = normalize_quaternion(q2, eps=eps)
dot_product = (q1_norm * q2_norm).sum(dim=-1, keepdim=True)
q2_corrected = torch.where(dot_product < 0, -q2_norm, q2_norm)
inner_product = (q1_norm * q2_corrected).sum(dim=-1)
# Clamp to valid range for arccos to avoid numerical issues
inner_product_clamped = torch.clamp(inner_product, min=-1.0 + eps, max=1.0 - eps)
theta = 2.0 * torch.arccos(torch.abs(inner_product_clamped))
return theta
def root_motion_reg(root_quats, root_pos):
return ((root_pos[1:] - root_pos[:-1])**2).mean(), (geodesic_loss(root_quats[1:], root_quats[:-1])**2).mean()
def joint_motion_coherence(quats_normed, parent_idx):
"""
Compute joint motion coherence loss to enforce smooth relative motion between parent-child joints.
"""
coherence_loss = 0
for j, parent in enumerate(parent_idx):
if parent != -1: # Skip root joint
parent_rot = quats_normed[:, parent] # (T, 4)
child_rot = quats_normed[:, j] # (T, 4)
# Compute relative rotation of child w.r.t. parent's local frame
# local_rot = parent_rot^(-1) * child_rot
local_rot = quat_multiply(quat_inverse(parent_rot), child_rot)
local_rot_velocity = local_rot[1:] - local_rot[:-1] # (T-1, 4)
coherence_loss += local_rot_velocity.pow(2).mean()
return coherence_loss
def read_flo_file(file_path):
"""
Read optical flow from .flo format file.
"""
with open(file_path, 'rb') as f:
magic = np.fromfile(f, np.float32, count=1)
if len(magic) == 0 or magic[0] != 202021.25:
raise ValueError(f'Invalid .flo file format: magic number {magic}')
w = np.fromfile(f, np.int32, count=1)[0]
h = np.fromfile(f, np.int32, count=1)[0]
data = np.fromfile(f, np.float32, count=2*w*h)
flow = data.reshape(h, w, 2)
return flow
def load_optical_flows(flow_dir, num_frames):
"""
Load sequence of optical flow files.
"""
flow_dir = Path(flow_dir)
flows = []
for i in range(num_frames - 1):
flow_path = flow_dir / f'flow_{i:04d}.flo'
if flow_path.exists():
flow = read_flo_file(flow_path)
flows.append(flow)
else:
raise ValueError("No flow files found")
return np.stack(flows, axis=0)
def rasterize_vertex_flow(flow_vertices, meshes, faces, image_size, renderer, eps = 1e-8):
"""
Rasterize per-vertex flow to dense flow field using barycentric interpolation.
"""
B, V, _ = flow_vertices.shape
device = flow_vertices.device
if isinstance(image_size, int):
H = W = image_size
else:
H, W = image_size
batch_meshes = join_meshes_as_batch([join_meshes_as_scene(m) for m in meshes]).to(device)
fragments = renderer.renderer.rasterizer(batch_meshes)
pix_to_face = fragments.pix_to_face # (B, H, W, K)
bary_coords = fragments.bary_coords # (B, H, W, K, 3)
flow_scene_list = []
for mesh_idx in range(B):
mesh = meshes[mesh_idx]
V_mesh = mesh.verts_packed().shape[0]
if V_mesh > flow_vertices.shape[1]:
raise ValueError(f"Mesh {mesh_idx} has {V_mesh} vertices but flow has {flow_vertices.shape[1]}")
flow_scene_list.append(flow_vertices[mesh_idx, :V_mesh])
flow_vertices_scene = torch.cat(flow_scene_list, dim=0).to(device)
faces_scene = batch_meshes.faces_packed()
flow_pred = torch.zeros(B, H, W, 2, device=device)
valid = pix_to_face[..., 0] >= 0
for b in range(B):
b_valid = valid[b] # (H,W)
if torch.count_nonzero(b_valid) == 0:
print(f"No valid pixels found for batch {b}")
continue
valid_indices = torch.nonzero(b_valid, as_tuple=True)
h_indices, w_indices = valid_indices
face_idxs = pix_to_face[b, h_indices, w_indices, 0] # (N,)
bary = bary_coords[b, h_indices, w_indices, 0] # (N,3)
max_face_idx = faces_scene.shape[0] - 1
if face_idxs.max() > max_face_idx:
raise RuntimeError(f"Face index {face_idxs.max()} exceeds max {max_face_idx}")
face_verts = faces_scene[face_idxs] # (N, 3)
f0, f1, f2 = face_verts.unbind(-1) # Each (N,)
max_vert_idx = flow_vertices_scene.shape[0] - 1
if max(f0.max(), f1.max(), f2.max()) > max_vert_idx:
raise RuntimeError(f"Vertex index exceeds flow_vertices_scene size {max_vert_idx}")
v0_flow = flow_vertices_scene[f0] # (N, 2)
v1_flow = flow_vertices_scene[f1] # (N, 2)
v2_flow = flow_vertices_scene[f2] # (N, 2)
# Interpolate using barycentric coordinates
b0, b1, b2 = bary.unbind(-1) # Each (N,)
# Ensure barycentric coordinates sum to 1 (numerical stability)
bary_sum = b0 + b1 + b2
b0 = b0 / (bary_sum + eps)
b1 = b1 / (bary_sum + eps)
b2 = b2 / (bary_sum + eps)
flow_interpolated = (
b0.unsqueeze(-1) * v0_flow +
b1.unsqueeze(-1) * v1_flow +
b2.unsqueeze(-1) * v2_flow
) # (N, 2)
# Update flow prediction
flow_pred[b, h_indices, w_indices] = flow_interpolated
return flow_pred
def calculate_flow_loss(flow_dir, device, mask, renderer, model):
"""
Calculate optical flow loss with improved error handling and flexibility.
"""
if device is None:
device = mask.device
T = mask.shape[0]
H, W = mask.shape[1:3]
if mask.shape[0] == T:
flow_mask = mask[1:] # Use frames 1 to T-1
else:
flow_mask = mask
flows_np = load_optical_flows(flow_dir, T)
flow_gt = torch.from_numpy(flows_np).float().to(device) # [T-1, H, W, 2]
vertices = model.deformed_vertices[0] # (T,V,3)
# Project vertices to get 2D flow
proj_t = renderer.project_points(vertices[:-1]) # (T-1,V,2) in pixels
proj_tp = renderer.project_points(vertices[1:])
vertex_flow = proj_tp - proj_t # (T-1,V,2) Δx,Δy
meshes = [model.get_mesh(t) for t in range(T)]
flow_pred = rasterize_vertex_flow(vertex_flow, meshes, model.faces[0], (H,W), renderer) # (B,H,W,2)
eps = 1e-3
diff = (flow_pred - flow_gt) * flow_mask.unsqueeze(-1) # (T-1, H, W, 2)
loss = torch.sqrt(diff.pow(2).sum(dim=-1) + eps**2) # Charbonnier loss
loss = loss.sum() / (flow_mask.sum() + 1e-6)
return loss
def normalize_depth_from_reference(depth_maps, reference_idx=0, invalid_value=-1.0, invert=False, eps = 1e-8):
"""
Normalize depth maps based on a reference frame with improved robustness.
"""
if depth_maps.dim() != 3:
raise ValueError(f"Expected depth_maps with 3 dimensions, got {depth_maps.dim()}")
T, H, W = depth_maps.shape
device = depth_maps.device
reference_depth = depth_maps[reference_idx]
valid_mask = (
(reference_depth != invalid_value) &
(reference_depth > 1e-8) & # Avoid very small positive values
torch.isfinite(reference_depth) # Exclude inf/nan
)
valid_values = reference_depth[valid_mask]
min_depth = torch.quantile(valid_values, 0.01) # 1st percentile
max_depth = torch.quantile(valid_values, 0.99) # 99th percentile
depth_range = max_depth - min_depth
if depth_range < eps:
logger.warning(f"Very small depth range ({depth_range:.6f}), using fallback normalization")
min_depth = valid_values.min()
max_depth = valid_values.max()
depth_range = max(max_depth - min_depth, eps)
scale = 1.0 / (max_depth - min_depth)
offset = -min_depth * scale
all_valid_mask = (
(depth_maps != invalid_value) &
(depth_maps > eps) &
torch.isfinite(depth_maps)
)
normalized_depths = torch.full_like(depth_maps, invalid_value)
if all_valid_mask.any():
normalized_values = depth_maps[all_valid_mask] * scale + offset
if invert:
normalized_values = 1.0 - normalized_values
normalized_depths[all_valid_mask] = normalized_values
return normalized_depths, scale.item(), offset.item()
def compute_depth_loss_normalized(mono_depths, zbuf_depths, mask):
"""
Compute normalized depth loss.
"""
device = zbuf_depths.device
# Normalize both depth types
zbuf_norm, z_scale, z_offset = normalize_depth_from_reference(zbuf_depths)
mono_norm, m_scale, m_offset = normalize_depth_from_reference(mono_depths, invert=True)
valid_zbuf = (zbuf_norm >= 0) & (zbuf_norm <= 1)
valid_mono = (mono_norm >= 0) & (mono_norm <= 1)
if mask.dtype != torch.bool:
mask = mask > 0.5
combined_mask = mask & valid_zbuf & valid_mono
num_valid = combined_mask.sum().item()
if num_valid == 0:
print("No valid pixels for depth loss computation")
return torch.tensor(0.0, device=device, requires_grad=True)
depth_diff = (zbuf_norm - mono_norm) * combined_mask.float()
loss = (depth_diff**2).sum() / num_valid
return loss |