File size: 19,470 Bytes
aae3ba1 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 |
import os
import cv2
import numpy as np
import torch
import matplotlib as mpl
from .video_utils import (
read_video_frames,
resize_frames_to_long_side,
save_to_video,
add_overlay_text
)
from typing import Optional, List, Tuple
from libs.models.mano_wrapper import MANO
from .render_utils import Renderer
class Config:
"""
Configuration class for file paths, parameters, and visual settings.
Paths are initialized with default values but can be overridden by arguments.
"""
def __init__(self, args=None):
# --- Paths (Overridden by CLI arguments) ---
self.VIDEO_ROOT = getattr(args, 'video_root', 'data/examples/videos')
self.LABEL_ROOT = getattr(args, 'label_root', 'data/examples/annotations')
self.SAVE_PATH = getattr(args, 'save_path', 'data/examples/visualize')
self.MANO_MODEL_PATH = getattr(args, 'mano_model_path', './weights/mano')
# --- Fixed Parameters ---
self.RENDER_SIZE_LONG_SIDE = 480
self.FPS = 15
# --- Fixed Colors and CMAPs ---
self.LEFT_CMAP = "inferno"
self.RIGHT_CMAP = "inferno"
# Base colors for the hands
self.LEFT_COLOR = np.array([0.6594, 0.6259, 0.7451])
self.RIGHT_COLOR = np.array([0.4078, 0.4980, 0.7451])
class HandVisualizer:
"""
Main class for loading data, configuring the renderer, and visualizing
the hand episode, including mesh and trajectory.
"""
def __init__(self, config: Config, render_gradual_traj: bool = False):
self.config = config
self.render_gradual_traj = render_gradual_traj
self.all_modes = ['cam', 'first']
if self.render_gradual_traj:
self.all_modes = ['cam', 'full', 'first']
# Initialize MANO and faces (use right hand model for both hands by default)
self.mano = MANO(model_path=self.config.MANO_MODEL_PATH).cuda()
faces_right = torch.from_numpy(self.mano.faces).float().cuda()
# MANO faces are defined for right hand, left hand faces need vertex order flip
self.faces_left = faces_right[:, [0, 2, 1]]
self.faces_right = faces_right
def _render_hand_trajectory(self, video_frames, hand_traj_wordspace, hand_mask, extrinsics, renderer: Renderer, mode: str):
"""
Renders hand mesh for one frame or hand trajectory across multiple frames,
depending on the mode ('cam', 'first', 'full').
"""
verts_left_worldspace, verts_right_worldspace = hand_traj_wordspace
left_hand_mask, right_hand_mask = hand_mask
R_w2c, t_w2c = extrinsics
num_total_frames = len(video_frames)
all_save_frames = []
# Determine rendering parameters based on mode
if mode == 'cam':
# Renders only the current frame's mesh
num_loop_frames = num_total_frames
# Single color for all frames
left_colors = self.config.LEFT_COLOR[np.newaxis, :].repeat(num_total_frames, axis=0)
right_colors = self.config.RIGHT_COLOR[np.newaxis, :].repeat(num_total_frames, axis=0)
elif mode == 'first':
# Renders the full trajectory onto the first frame
num_loop_frames = 1
left_colors = self.config.LEFT_COLOR[np.newaxis, :].repeat(num_total_frames, axis=0)
right_colors = self.config.RIGHT_COLOR[np.newaxis, :].repeat(num_total_frames, axis=0)
elif mode == 'full':
# Renders a gradual trajectory for each frame
num_loop_frames = num_total_frames
# Generate color sequence for trajectory
left_colors, right_colors = generate_hand_colors(num_total_frames, self.config.LEFT_CMAP, self.config.RIGHT_CMAP)
else:
raise ValueError(f'Unknown rendering mode: {mode}')
for current_frame_idx in range(num_loop_frames):
if not mode == 'first':
print(f'Processing frame {current_frame_idx + 1}/{num_loop_frames}', end='\r')
# Start with the base video frame (copied and normalized)
curr_img_overlay = video_frames[current_frame_idx].copy().astype(np.float32) / 255.0
# Calculate camera space vertices for all frames relative to the *current* camera pose
R_w2c_cur = R_w2c[current_frame_idx]
t_w2c_cur = t_w2c[current_frame_idx]
# R_w2c_cur: (3, 3), t_w2c_cur: (3, 1). Need to broadcast to all frames (T)
verts_left_camspace = (
R_w2c_cur @ verts_left_worldspace.transpose(0, 2, 1) + t_w2c_cur
).transpose(0, 2, 1)
verts_right_camspace = (
R_w2c_cur @ verts_right_worldspace.transpose(0, 2, 1) + t_w2c_cur
).transpose(0, 2, 1)
# Determine the segment of the trajectory to render for the current frame
if mode == 'cam':
# Render only the current frame's mesh (from index to index)
start_traj_idx = current_frame_idx
end_traj_idx = current_frame_idx + 1
transparency = [1.0]
elif mode == 'first':
# Render full trajectory on frame 0 (index 0 to T)
start_traj_idx = 0
end_traj_idx = num_total_frames
transparency = [1.0] * (end_traj_idx - start_traj_idx)
# The loop only runs once for mode='first'
if current_frame_idx > 0: continue
elif mode == 'full':
# Render gradual trajectory (from index to T)
start_traj_idx = current_frame_idx
end_traj_idx = num_total_frames
# Gradual transparency for the trajectory: older points are more transparent
transparency = np.linspace(0.4, 0.7, end_traj_idx - start_traj_idx)
else:
raise ValueError(f'Unknown rendering mode: {mode}')
# Iterate over the trajectory segment
for traj_idx, kk in enumerate(range(start_traj_idx, end_traj_idx)):
if mode == 'first':
print(f'Processing frame {traj_idx + 1}/{num_total_frames}', end='\r')
curr_img_overlay = video_frames[current_frame_idx].copy().astype(np.float32)/255
# Get hand data for the trajectory point 'kk'
left_mask_k = left_hand_mask[kk]
right_mask_k = right_hand_mask[kk]
transp_k = transparency[traj_idx] if len(transparency) > traj_idx else 1.0
left_verts_list, left_color_list, left_face_list = ([], [], [])
right_verts_list, right_color_list, right_face_list = ([], [], [])
if left_mask_k != 0:
left_verts_list = [torch.from_numpy(verts_left_camspace[kk]).float().cuda()]
# Repeat color for all 778 vertices
left_color_list = [torch.from_numpy(left_colors[kk]).float().unsqueeze(0).repeat(778, 1).cuda()]
left_face_list = [self.faces_left]
if right_mask_k != 0:
right_verts_list = [torch.from_numpy(verts_right_camspace[kk]).float().cuda()]
right_color_list = [torch.from_numpy(right_colors[kk]).float().unsqueeze(0).repeat(778, 1).cuda()]
right_face_list = [self.faces_right]
verts_list = left_verts_list + right_verts_list
faces_list = left_face_list + right_face_list
colors_list = left_color_list + right_color_list
if verts_list:
# Render the mesh onto the current image
rend, mask = renderer.render(verts_list, faces_list, colors_list)
rend = rend[..., ::-1] # RGB to BGR
color_mesh = rend.astype(np.float32) / 255.0
valid_mask = mask[..., None].astype(np.float32)
# Alpha blending for the mesh overlay:
# new_image = base * (1-mask) + mesh * mask * alpha + base * mask * (1-alpha)
curr_img_overlay = (
curr_img_overlay[:, :, :3] * (1 - valid_mask) +
color_mesh[:, :, :3] * valid_mask * transp_k +
curr_img_overlay[:, :, :3] * valid_mask * (1 - transp_k)
)
if mode == 'first':
# Finalize image format and color space
final_frame = (curr_img_overlay * 255).astype(np.uint8)
final_frame = cv2.cvtColor(final_frame, cv2.COLOR_BGR2RGB)
all_save_frames.append(final_frame)
if mode == 'cam' or mode == 'full':
# Finalize image format and color space
final_frame = (curr_img_overlay * 255).astype(np.uint8)
final_frame = cv2.cvtColor(final_frame, cv2.COLOR_BGR2RGB)
all_save_frames.append(final_frame)
print(f'Finished rendering with mode: {mode}')
return all_save_frames
def process_episode(self, episode_name: str):
"""Loads data and orchestrates the visualization process for a single episode."""
print(f'\nProcessing episode: {episode_name}')
# 1. Load Paths and Check Existence
dataset_name = episode_name.split('_')[0]
ep_name = episode_name.split('_')[-2] + '_' + episode_name.split('_')[-1]
video_name = episode_name.replace(f'{dataset_name}_', '').replace(f'_{ep_name}', '')
video_path = os.path.join(self.config.VIDEO_ROOT, f'{video_name}.mp4')
label_path = os.path.join(self.config.LABEL_ROOT, episode_name + '.npy')
if not os.path.exists(label_path):
print(f'Episode file {label_path} does not exist, skipping...')
return
# 2. Load Episode Data
cap = cv2.VideoCapture(video_path)
episode_info = np.load(label_path, allow_pickle=True).item()
start_frame, end_frame = get_frame_interval(episode_info)
R_w2c, t_w2c, normalized_intrinsics = get_camera_info(episode_info)
caption_left, caption_right, hand_type = get_caption_info(episode_info)
(verts_left_worldspace, left_hand_mask), (verts_right_worldspace, right_hand_mask) = \
get_hand_labels(episode_info, self.mano)
# 3. Read and Resize Video Frames
video_frames = read_video_frames(cap, start_frame=start_frame, end_frame=end_frame, interval=1)
resize_video_frames = resize_frames_to_long_side(video_frames, self.config.RENDER_SIZE_LONG_SIDE)
H, W, _ = resize_video_frames[0].shape
# 4. Initialize Renderer
# Denormalize intrinsics based on the new frame size (W, H)
intrinsics_denorm = normalized_intrinsics.copy()
intrinsics_denorm[0] *= W
intrinsics_denorm[1] *= H
fx_exo = intrinsics_denorm[0, 0]
fy_exo = intrinsics_denorm[1, 1]
renderer = Renderer(W, H, (fx_exo, fy_exo), 'cuda')
# 5. Render Hands for All Modes
all_rendered_frames = []
hand_traj_wordspace = (verts_left_worldspace, verts_right_worldspace)
hand_mask = (left_hand_mask, right_hand_mask)
extrinsics = (R_w2c, t_w2c)
for mode in self.all_modes:
save_frames = self._render_hand_trajectory(
resize_video_frames,
hand_traj_wordspace,
hand_mask,
extrinsics,
renderer,
mode=mode
)
all_rendered_frames.append(save_frames)
# 6. Concatenate Frames and Add Captions
final_save_frames = []
num_frames = len(all_rendered_frames[0])
# Select primary caption and extract opposite intervals for text lookup
caption_primary = caption_right if hand_type == 'right' else caption_left
caption_opposite = caption_left if hand_type == 'right' else caption_right
opposite_intervals = [interval for _, interval in caption_opposite]
for frame_idx in range(num_frames):
# Concatenate frames from different modes side by side
curr_img_overlay = np.concatenate(
[all_rendered_frames[mode_idx][frame_idx] for mode_idx in range(len(self.all_modes))],
axis=1
)
# Get caption for the primary hand (assumes primary caption only has one interval: [0])
overlay_text_primary = caption_primary[0][0]
# Get caption for the opposite hand based on the current frame index
opposite_idx = find_caption_index(frame_idx, opposite_intervals)
overlay_text_opposite = caption_opposite[opposite_idx][0] if opposite_idx is not None else 'None.'
# Format and add the full overlay text
overlay_text_full = generate_overlay_text(
overlay_text_primary,
overlay_text_opposite,
hand_type
)
add_overlay_text(curr_img_overlay, overlay_text_full)
final_save_frames.append(curr_img_overlay)
# 7. Save Final Video
os.makedirs(self.config.SAVE_PATH, exist_ok=True)
save_to_video(final_save_frames, f'{self.config.SAVE_PATH}/{episode_name}.mp4', fps=self.config.FPS)
print(f'\nSuccessfully saved episode to {self.config.SAVE_PATH}/{episode_name}.mp4')
def find_caption_index(frame_index: int, intervals: list[tuple[int, int]]) -> Optional[int]:
"""Finds the interval index for a given frame index."""
for idx, (start, end) in enumerate(intervals):
if start <= frame_index <= end:
return idx
return None
def generate_hand_colors(T: int, left_cmap: str, right_cmap: str) -> tuple[np.ndarray, np.ndarray]:
"""
Generates RGB color sequences for left and right hands over T frames.
Returns colors in shape (T, 3), normalized 0-1, based on the specified colormaps.
"""
t_norm = np.linspace(0, 0.95, T)
left_colors = mpl.colormaps.get_cmap(left_cmap)(t_norm)[:, :3]
right_colors = mpl.colormaps.get_cmap(right_cmap)(t_norm)[:, :3]
return left_colors, right_colors
def get_frame_interval(episode_info: dict) -> tuple[int, int]:
"""Extracts start (inclusive) and end (exclusive) frame indices from episode info."""
video_decode_frames = episode_info['video_decode_frame']
start_frame = video_decode_frames[0]
end_frame = video_decode_frames[-1] + 1
return start_frame, end_frame
def normalize_camera_intrinsics(intrinsics: np.ndarray) -> np.ndarray:
"""
Normalizes intrinsics based on the assumption that the principal point
is at the image center (image size is 2*cx, 2*cy).
"""
# Create a deep copy to avoid modifying the original array
normalized_intrinsics = intrinsics.copy()
normalized_intrinsics[0] /= normalized_intrinsics[0, 2] * 2
normalized_intrinsics[1] /= normalized_intrinsics[1, 2] * 2
return normalized_intrinsics
def get_camera_info(episode_info: dict) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
"""
Extracts and normalizes camera intrinsics and extrinsics (world-to-cam).
"""
extrinsics = episode_info['extrinsics'] # world2cam, shape (T, 4, 4)
R_w2c = extrinsics[:, :3, :3].copy()
t_w2c = extrinsics[:, :3, 3:].copy() # shape (T, 3, 1)
intrinsics = episode_info['intrinsics'].copy()
normalized_intrinsics = normalize_camera_intrinsics(intrinsics)
return R_w2c, t_w2c, normalized_intrinsics
def get_caption_info(episode_info: dict) -> tuple[list, list, str]:
"""
Extracts and formats caption information for left and right hands.
Adds a large interval if captions are empty to cover all frames.
"""
hand_type = episode_info['anno_type']
caption_right = episode_info['text'].get('right', [])
caption_left = episode_info['text'].get('left', [])
# Ensure captions are not empty to simplify downstream logic
if not caption_right:
caption_right = [['None.', (0, 10000)]] # large interval to cover all frames
if not caption_left:
caption_left = [['None.', (0, 10000)]]
return caption_left, caption_right, hand_type
def get_hand_labels(episode_info: dict, mano: MANO):
"""
Processes hand labels (pose, shape, translation, orientation) through the MANO model
to obtain hand vertices in world space.
"""
left_labels = episode_info['left']
right_labels = episode_info['right']
# --- Left Hand Processing ---
left_hand_mask = left_labels['kept_frames']
verts_left, _ = process_single_hand_labels(left_labels, left_hand_mask, mano, is_left=True)
# --- Right Hand Processing ---
right_hand_mask = right_labels['kept_frames']
verts_right, _ = process_single_hand_labels(right_labels, right_hand_mask, mano)
return (verts_left, left_hand_mask), (verts_right, right_hand_mask)
def process_single_hand_labels(hand_labels: dict, hand_mask: np.ndarray, mano: MANO, is_left: bool = False):
"""
Helper function to compute MANO vertices for a single hand (left or right).
"""
T = len(hand_mask)
wrist_worldspace = hand_labels['transl_worldspace'].reshape(-1, 1, 3)
wrist_orientation = hand_labels['global_orient_worldspace']
beta = hand_labels['beta']
pose = hand_labels['hand_pose']
# Set pose to identity for masked frames (no hand present)
identity = np.eye(3, dtype=pose.dtype)
identity_block = np.broadcast_to(identity, (pose.shape[1], 3, 3))
mask_indices = (hand_mask == 0)
if np.any(mask_indices):
pose[mask_indices] = identity_block
# pose[hand_mask == 0] = identity_block
beta_torch = torch.from_numpy(beta).float().cuda().unsqueeze(0).repeat(T, 1)
pose_torch = torch.from_numpy(pose).float().cuda()
# Placeholder for global orientation in MANO forward pass (will be applied manually later)
global_rot_placeholder = torch.eye(3).float().unsqueeze(0).unsqueeze(0).cuda().repeat(T, 1, 1, 1)
# MANO forward pass
mano_out = mano(betas=beta_torch, hand_pose=pose_torch, global_orient=global_rot_placeholder)
verts = mano_out.vertices.cpu().numpy()
joints = mano_out.joints.cpu().numpy()
# Apply the wrist orientation and translation to get world space coordinates
# X-axis flip for the left hand.
if is_left:
verts[:, :, 0] *= -1
joints[:, :, 0] *= -1
# World space transformation: R @ (V - J0) + T
# (T, 778, 3) = (T, 3, 3) @ (T, 3, 778) + (T, 3, 1) -> (T, 3, 778) -> (T, 778, 3)
verts_worldspace = (
wrist_orientation @
(verts - joints[:, 0][:, None]).transpose(0, 2, 1)
).transpose(0, 2, 1) + wrist_worldspace
return verts_worldspace, joints[:, 0]
def generate_overlay_text(overlay_text: str, overlay_text_opposite: str, hand_type: str) -> str:
"""Formats the caption string based on the primary hand type."""
if hand_type == 'right':
return f'Left: {overlay_text_opposite} | Right: {overlay_text}'
else:
return f'Left: {overlay_text} | Right: {overlay_text_opposite}' |