File size: 2,488 Bytes
4c79aec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import numpy as np


def resize_or_pad_sequence(arr: np.ndarray, target_frames: int) -> np.ndarray:
    if arr.shape[0] == target_frames:
        return arr
    if arr.shape[0] > target_frames:
        idx = np.linspace(0, arr.shape[0] - 1, target_frames).round().astype(int)
        return arr[idx]
    pad = np.repeat(arr[-1:], target_frames - arr.shape[0], axis=0)
    return np.concatenate([arr, pad], axis=0)


def normalize_pose_sequence(pose: np.ndarray, min_conf: float = 0.05) -> np.ndarray:
    pose = pose.astype(np.float32).copy()
    xy = pose[..., :2]
    conf = pose[..., 2:3] if pose.shape[-1] > 2 else np.ones((*pose.shape[:2], 1), dtype=np.float32)
    valid = conf[..., 0] > min_conf
    out_xy = np.zeros_like(xy, dtype=np.float32)
    for t in range(pose.shape[0]):
        mask = valid[t]
        if mask.sum() < 2:
            continue
        pts = xy[t, mask]
        center = (pts.min(axis=0) + pts.max(axis=0)) / 2.0
        size = np.maximum(pts.max(axis=0) - pts.min(axis=0), 1.0)
        scale = float(max(size[0], size[1], 1.0))
        out_xy[t] = (xy[t] - center) / scale
    return np.concatenate([out_xy, conf.astype(np.float32)], axis=-1)


def add_keypoint_noise(pose: np.ndarray, sigma: float, rng: np.random.Generator) -> np.ndarray:
    if sigma <= 0:
        return pose
    noisy = pose.copy()
    noisy[..., :2] += rng.normal(0.0, sigma, size=noisy[..., :2].shape).astype(np.float32)
    return noisy


def apply_frame_drop(pose: np.ndarray, drop_ratio: float, rng: np.random.Generator) -> np.ndarray:
    if drop_ratio <= 0:
        return pose
    out = pose.copy()
    frames = out.shape[0]
    keep = rng.random(frames) > drop_ratio
    keep[0] = True
    last = out[0].copy()
    for t in range(frames):
        if keep[t]:
            last = out[t].copy()
        else:
            out[t] = last
    return out


def pose_to_feature_vector(
    pose: np.ndarray,
    use_confidence: bool = True,
    use_velocity: bool = True,
    min_conf: float = 0.05,
) -> np.ndarray:
    norm = normalize_pose_sequence(pose, min_conf=min_conf)
    xy = norm[..., :2]
    parts = [xy.reshape(xy.shape[0], -1)]
    if use_confidence:
        parts.append(norm[..., 2:3].reshape(norm.shape[0], -1))
    if use_velocity:
        vel = np.zeros_like(xy, dtype=np.float32)
        vel[1:] = xy[1:] - xy[:-1]
        parts.append(vel.reshape(vel.shape[0], -1))
    return np.concatenate(parts, axis=1).astype(np.float32)