Files
AV-Live/data_only_viz/hand_features.py
T
2026-06-30 18:22:38 +02:00

109 lines
4.2 KiB
Python

"""Turn MediaPipe hand landmarks into a compact expressive feature vector.
Mirrors the kinematics pattern of action_head.FeatureExtractor: a small ring
buffer per hand slot, finite-guarded, clamped to declared ranges. Output feeds
both the OSC /pose/hands route (audio) and the Metal renderer uniforms (visual).
"""
from __future__ import annotations
import math
from collections import deque
WRIST, THUMB_TIP, MIDDLE_MCP, PINKY_TIP = 0, 4, 9, 20
# index/middle/ring/pinky tips + their MCP knuckles, for the "all fingers
# curled" fist metric (thumb excluded — a fist curls the four fingers).
_FINGER_TIPS = (8, 12, 16, 20)
_FINGER_MCPS = (5, 9, 13, 17)
NEUTRAL_HAND = {"cx": 0.5, "cy": 0.5, "openness": 0.0, "speed": 0.0, "fist": 0.0}
# openness calibration: (span/size) maps fist~0.3 -> 0, open~2.0 -> 1
_OPEN_LO, _OPEN_HI = 0.5, 2.0
# fist calibration: per-finger tip->MCP distance / hand size. curled ~0.25,
# extended ~0.75. closedness = 1 at/below LO, 0 at/above HI.
_EXT_LO, _EXT_HI = 0.25, 0.75
def _finite(v: float, fallback: float) -> float:
return v if isinstance(v, float) and math.isfinite(v) else fallback
def _coord(p, attr: str, idx: int, fallback: float = 0.5) -> float:
"""Extract a coordinate from a landmark object or an indexable row."""
if hasattr(p, attr):
return float(getattr(p, attr))
try:
return float(p[idx])
except (TypeError, IndexError):
return fallback
def _clamp(v: float, lo: float, hi: float) -> float:
return lo if v < lo else hi if v > hi else v
class HandFeatureExtractor:
def __init__(self, history: int = 5):
self._history = max(2, history)
# two screen slots: 0 = leftmost (L), 1 = rightmost (R)
self._buf = [deque(maxlen=self._history), deque(maxlen=self._history)]
def _features(self, lm: list) -> dict:
xs = [_finite(_coord(p, "x", 0), 0.5) for p in lm[:21]]
ys = [_finite(_coord(p, "y", 1), 0.5) for p in lm[:21]]
cx = _clamp(sum(xs) / len(xs), 0.0, 1.0)
cy = _clamp(sum(ys) / len(ys), 0.0, 1.0)
size = math.hypot(xs[MIDDLE_MCP] - xs[WRIST], ys[MIDDLE_MCP] - ys[WRIST])
size = size if size > 1e-4 else 1e-4
span = math.hypot(xs[THUMB_TIP] - xs[PINKY_TIP],
ys[THUMB_TIP] - ys[PINKY_TIP])
openness = _clamp((span / size - _OPEN_LO) / (_OPEN_HI - _OPEN_LO),
0.0, 1.0)
# fist = 1.0 only when ALL four fingers are curled to the palm (min over
# fingers): one extended finger drops it toward 0, so a true fist is
# required rather than just a small thumb-pinky span.
fist = 1.0
for tip, mcp in zip(_FINGER_TIPS, _FINGER_MCPS):
ext = math.hypot(xs[tip] - xs[mcp], ys[tip] - ys[mcp]) / size
curled = _clamp(1.0 - (ext - _EXT_LO) / (_EXT_HI - _EXT_LO), 0.0, 1.0)
fist = min(fist, curled)
return {"cx": cx, "cy": cy, "openness": openness, "fist": fist}
def _speed(self, slot: int, cx: float, cy: float) -> float:
buf = self._buf[slot]
spd = 0.0
if buf:
px, py = buf[-1]
spd = math.hypot(cx - px, cy - py)
buf.append((cx, cy))
return _clamp(spd, 0.0, 1.0)
def step(self, hands: list) -> dict:
feats = []
for lm in hands:
try:
valid = lm is not None and len(lm) >= 21
except TypeError:
continue
if valid:
feats.append(self._features(lm))
if not feats:
self._buf[0].clear()
self._buf[1].clear()
return {"L": None, "R": None, "dist": 0.0}
feats.sort(key=lambda f: f["cx"]) # leftmost first
left = feats[0]
right = feats[-1] if len(feats) > 1 else None
left["speed"] = self._speed(0, left["cx"], left["cy"])
out_l = left
out_r = None
dist = 0.0
if right is not None:
right["speed"] = self._speed(1, right["cx"], right["cy"])
out_r = right
dist = _clamp(math.hypot(right["cx"] - left["cx"],
right["cy"] - left["cy"]), 0.0, 1.0)
else:
self._buf[1].clear()
return {"L": out_l, "R": out_r, "dist": dist}