File size: 6,021 Bytes
3b18ebd | 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 | # -*- coding: utf-8 -*-
"""55 節點骨架圖定義（ST-GCN 用）。
節點順序與 extract_features.py 完全一致：
0..11 pose，對應 MediaPipe 的 [11,12,13,14,15,16,17,18,19,20,21,22]
12..32 左手 21 點
33..53 右手 21 點
54 虛擬頸點（雙肩中點）
pose 那 12 點的實際意義：
slot 0=左肩(11) 1=右肩(12) 2=左肘(13) 3=右肘(14) 4=左腕(15) 5=右腕(16)
slot 6=左小指(17) 7=右小指(18) 8=左食指(19) 9=右食指(20) 10=左拇指(21) 11=右拇指(22)
"""
from __future__ import annotations
import numpy as np
import config as C
# ---------------------------------------------------------------- 節點編號
L_SHOULDER, R_SHOULDER = 0, 1
L_ELBOW, R_ELBOW = 2, 3
L_WRIST, R_WRIST = 4, 5
L_PINKY, R_PINKY = 6, 7
L_INDEX, R_INDEX = 8, 9
L_THUMB, R_THUMB = 10, 11
NECK = C.CENTER_IDX # 54
LH0 = C.LHAND_SLICE.start # 12，左手腕
RH0 = C.RHAND_SLICE.start # 33，右手腕
# ---------------------------------------------------------------- 邊
# 上半身骨架
POSE_EDGES = [
(NECK, L_SHOULDER), (NECK, R_SHOULDER),
(L_SHOULDER, L_ELBOW), (L_ELBOW, L_WRIST),
(R_SHOULDER, R_ELBOW), (R_ELBOW, R_WRIST),
(L_WRIST, L_PINKY), (L_WRIST, L_INDEX), (L_WRIST, L_THUMB),
(R_WRIST, R_PINKY), (R_WRIST, R_INDEX), (R_WRIST, R_THUMB),
]
# MediaPipe 單手 21 點的標準連接
HAND_EDGES = [
(0, 1), (1, 2), (2, 3), (3, 4), # 拇指
(0, 5), (5, 6), (6, 7), (7, 8), # 食指
(5, 9), (9, 10), (10, 11), (11, 12), # 中指
(9, 13), (13, 14), (14, 15), (15, 16), # 無名指
(13, 17), (17, 18), (18, 19), (19, 20), # 小指
(0, 17), # 掌根
]
def build_edges():
edges = list(POSE_EDGES)
for off in (LH0, RH0):
edges += [(a + off, b + off) for a, b in HAND_EDGES]
# 把 pose 的手腕接到手部骨架的手腕，讓手臂和手掌是連通的
edges += [(L_WRIST, LH0), (R_WRIST, RH0)]
return edges
# ---------------------------------------------------------------- 鄰接矩陣
def _hop_distance(num_nodes: int, edges, max_hop: int = 1) -> np.ndarray:
A = np.zeros((num_nodes, num_nodes))
for i, j in edges:
A[i, j] = 1
A[j, i] = 1
hop = np.full((num_nodes, num_nodes), np.inf)
powers = [np.linalg.matrix_power(A, d) for d in range(max_hop + 1)]
arrive = (np.stack(powers) > 0)
for d in range(max_hop, -1, -1):
hop[arrive[d]] = d
return hop
def _normalize(A: np.ndarray) -> np.ndarray:
"""D^-1 A，避免度數高的節點主導。"""
deg = A.sum(axis=0)
Dinv = np.zeros_like(A)
nz = deg > 0
Dinv[nz, nz] = deg[nz] ** -1
return A @ Dinv
def build_adjacency(num_nodes: int = C.NUM_POINTS, strategy: str = "spatial") -> np.ndarray:
"""回傳 (K, V, V) 的鄰接矩陣堆疊。
spatial（ST-GCN 原論文的分割方式）分成三組：
0 根節點：自己 + 與自己離頸點等距的鄰居
1 向心：比自己更靠近頸點的鄰居（代表軀幹方向的運動）
2 離心：比自己更遠離頸點的鄰居（代表末端手指的運動）
對手語很合理——同一個手勢，重點常在末端相對於軀幹怎麼動。
"""
edges = build_edges()
hop = _hop_distance(num_nodes, edges, max_hop=1)
adjacency = (hop <= 1).astype(float) # 含自環
norm = _normalize(adjacency)
dist_to_center = _hop_distance(num_nodes, edges, max_hop=num_nodes)[NECK]
if strategy == "uniform":
return norm[None, ...]
if strategy != "spatial":
raise ValueError(f"未知的 strategy: {strategy}")
root = np.zeros_like(norm)
close = np.zeros_like(norm)
far = np.zeros_like(norm)
for i in range(num_nodes):
for j in range(num_nodes):
if hop[j, i] != 1 and i != j:
continue
di, dj = dist_to_center[i], dist_to_center[j]
if dj == di:
root[j, i] = norm[j, i]
elif dj > di:
far[j, i] = norm[j, i]
else:
close[j, i] = norm[j, i]
return np.stack([root, close, far])
def build_parents(num_nodes: int = C.NUM_POINTS):
"""以頸點為根做 BFS，回傳 (parents, bfs_order)。
parents[v] = v 的父節點索引，根節點為 -1。
掌根那圈迴路在 BFS 時自然被展開成樹，不影響。
骨向量與肢段縮放都需要這個樹狀結構。
"""
adj = [[] for _ in range(num_nodes)]
for i, j in build_edges():
adj[i].append(j)
adj[j].append(i)
parents = np.full(num_nodes, -1, dtype=int)
visited = np.zeros(num_nodes, dtype=bool)
order, queue = [], [NECK]
visited[NECK] = True
while queue:
v = queue.pop(0)
order.append(v)
for u in adj[v]:
if not visited[u]:
visited[u] = True
parents[u] = v
queue.append(u)
return parents, order
def sanity_check():
"""檢查圖是連通的、沒有孤立節點。"""
edges = build_edges()
deg = np.zeros(C.NUM_POINTS, dtype=int)
for i, j in edges:
deg[i] += 1
deg[j] += 1
isolated = np.where(deg == 0)[0]
hop = _hop_distance(C.NUM_POINTS, edges, max_hop=C.NUM_POINTS)
unreachable = np.where(~np.isfinite(hop[NECK]))[0]
return {
"num_edges": len(edges),
"isolated_nodes": isolated.tolist(),
"unreachable_from_neck": unreachable.tolist(),
"max_hop_from_neck": float(np.nanmax(hop[NECK][np.isfinite(hop[NECK])])),
}
if __name__ == "__main__":
print(sanity_check())
A = build_adjacency()
print("鄰接矩陣 shape:", A.shape, " 每組非零數:", [int((a > 0).sum()) for a in A])
|