Spaces:
Sleeping
Sleeping
File size: 8,000 Bytes
0d0412d | 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 | """
Utility Functions for Pose Processing
======================================
Core utility functions supporting video preprocessing and pose normalization
across both the pose and hybrid pipelines. Handles geometric transformations,
video segment extraction, and crop configuration resolution.
Key Components:
1. Pose Normalization (normalize_pose, normalize_sequence)
- Person-centric coordinate system transformation
- Three-step process: Centering → Alignment → Scaling
- Hip-centered with spine-aligned Y-axis
- Scale-invariant via spine length normalization
2. Video Segment Extraction (get_segment_bounds)
- Extracts relevant portions of shot videos
- Supports tail-based, middle-based, or full extraction
- Configurable via segment_rules in params.yaml
3. Crop Configuration (resolve_crop_config_for_video)
- Per-video crop overrides based on video number
- Supports wildcard patterns for batch configuration
- Handles pre-cropped files with "(N)" naming pattern
4. File Pattern Detection (should_skip_crop, extract_video_number)
- Detects pre-cropped files that should skip cropping
- Extracts video numbers for per-video configuration
Normalization Algorithm:
1. Center pose at hip midpoint (joints 23, 24)
2. Align Y-axis with spine direction (hip → shoulder center)
3. Compute orthogonal X-axis from shoulder vector
4. Z-axis via cross product for right-handed system
5. Scale by spine length for size invariance
Dependencies:
External: numpy, re
Configuration (params.yaml):
segment_rules: Video segment extraction settings
crop_overrides: Per-video crop configuration
Author: IPD Research Team
Version: 1.0.0
"""
import numpy as np
import re
def normalize_pose(keypoints_3d):
"""
Normalizes a single 3D pose (33, 3) to a standard, person-centric coordinate system.
(Centering -> Alignment -> Scaling)
"""
if keypoints_3d.shape != (33, 3): return keypoints_3d
LEFT_SHOULDER, RIGHT_SHOULDER = 11, 12
LEFT_HIP, RIGHT_HIP = 23, 24
# 1. Centering
hip_center = (keypoints_3d[LEFT_HIP] + keypoints_3d[RIGHT_HIP]) / 2.0
centered = keypoints_3d - hip_center
# 2. Alignment
shoulder_center = (centered[LEFT_SHOULDER] + centered[RIGHT_SHOULDER]) / 2.0
spine_len = np.linalg.norm(shoulder_center)
if spine_len < 1e-6: return centered
new_y = shoulder_center / spine_len
right_shoulder_vec = centered[RIGHT_SHOULDER] - centered[LEFT_SHOULDER]
proj = np.dot(right_shoulder_vec, new_y) * new_y
new_x = right_shoulder_vec - proj
if np.linalg.norm(new_x) < 1e-6:
new_x = np.cross(new_y, [0, 1, 0]) if abs(new_y[0]) > 0.5 else np.cross(new_y, [1, 0, 0])
new_x /= np.linalg.norm(new_x)
new_z = np.cross(new_x, new_y)
rotation = np.array([new_x, new_y, new_z])
aligned = np.dot(centered, rotation.T)
# 3. Scaling
return aligned / spine_len
def normalize_sequence(keypoints_sequence):
"""Applies pose normalization to an entire sequence of frames."""
return np.array([normalize_pose(frame) for frame in keypoints_sequence])
def should_skip_crop(video_path_or_name: str) -> bool:
"""Return True if the filename matches the pattern "name (N).ext".
This matches filenames that contain a space followed by a parenthesized
integer index before the extension, e.g. "backhand_drive (1).mp4".
The check accepts either a full path or a bare filename.
"""
import re, os
name = os.path.basename(video_path_or_name)
# Match 'something (123).ext' where the number is one or more digits
return re.search(r"\s\(\d+\)\.[^.]+$", name) is not None
def extract_video_number(video_path_or_name: str) -> int | None:
"""Extract the leading numeric id from a filename/path.
Examples:
- "001.mp4" -> 1
- "006_win_3.npz" -> 6
- "12_some_name.mov" -> 12
Returns None if no leading number exists.
"""
import os
import re
name = os.path.basename(video_path_or_name)
m = re.match(r"^(\d+)", name)
if not m:
return None
try:
return int(m.group(1))
except Exception:
return None
def resolve_crop_config_for_video(
video_path: str,
base_crop_config: dict,
crop_overrides: dict | None = None,
) -> dict:
"""Return crop_config with per-shot per-video overrides applied.
- Only overrides keys present in the matching rule (currently just 'bottom').
- Matches shot name by checking path components against crop_overrides keys.
- Matches by leading video number in filename.
- If multiple ranges match, the last matching range wins.
"""
import copy
import os
effective = copy.deepcopy(base_crop_config) if base_crop_config is not None else {
'top': 0.0,
'bottom': 0.0,
'left': 0.0,
'right': 0.0,
}
if not crop_overrides:
return effective
video_num = extract_video_number(video_path)
if video_num is None:
return effective
comps = set(os.path.normpath(video_path).split(os.path.sep))
matched_shot = None
for shot in crop_overrides.keys():
if shot in comps:
matched_shot = shot
break
if not matched_shot:
return effective
rules = crop_overrides.get(matched_shot) or []
for rule in rules:
try:
start = int(rule.get('start'))
end = int(rule.get('end'))
except Exception:
continue
if start <= video_num <= end:
for k, v in rule.items():
if k in {'start', 'end'}:
continue
effective[k] = v
return effective
def get_tail_seconds_for_video(video_path: str, default: float = 1.75) -> float:
"""Return the tail window length in seconds for a given video path.
Special-case: use 2.0 seconds for forehand_lift and forehand_clear shots.
The check looks at any path component equal to those shot names so it
works with full paths or relative paths.
"""
import os
p = os.path.normpath(video_path)
comps = p.split(os.path.sep)
shots_2s = {'forehand_lift', 'forehand_clear'}
if any(c in shots_2s for c in comps):
return 2.0
return default
def get_segment_bounds(
video_path: str,
fps: float,
total_frames: int,
default_seconds: float = 1.75,
segment_cfg: dict | None = None,
):
"""Return (start_frame, frame_count) for the segment to process for a given video.
Config-driven rules (segment_cfg):
- default_seconds: fallback window length (tail).
- tail_seconds: window length for tail_shots.
- tail_shots: list of shot folder names to use tail_seconds.
- middle_shots: mapping of shot folder name -> seconds for middle window.
If no config is provided, uses hardcoded defaults (tail, 1.75s).
"""
import os
if segment_cfg is None:
segment_cfg = {}
default_seconds = segment_cfg.get('default_seconds', default_seconds)
tail_seconds = segment_cfg.get('tail_seconds', default_seconds)
tail_shots = set(segment_cfg.get('tail_shots', []))
middle_shots = segment_cfg.get('middle_shots', {})
p = os.path.normpath(video_path)
comps = p.split(os.path.sep)
# Middle-shot rule
for shot, secs in middle_shots.items():
if shot in comps:
frames = int(float(secs) * fps)
center = total_frames // 2
start = max(0, center - frames // 2)
return start, frames
# Tail-shot rule
if any(shot in comps for shot in tail_shots):
frames = int(float(tail_seconds) * fps)
start = max(0, total_frames - frames)
return start, frames
# Default tail
frames = int(float(default_seconds) * fps)
start = max(0, total_frames - frames)
return start, frames |