Spaces:
Running
Running
File size: 6,051 Bytes
d646dd7 95f5290 d646dd7 95f5290 d646dd7 95f5290 d646dd7 | 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 | """Local pose backends used by the bundled black-glove detector.
This is the self-contained equivalent of the hand_3d_d435 pose backend
contract. It deliberately resolves the RTMPose model and MediaPipe task
from the runtime/package itself; it never imports the legacy tools directory.
"""
from __future__ import annotations
import os
from typing import Optional, Tuple
import cv2
import numpy as np
_PACKAGE_ROOT = os.path.dirname(os.path.abspath(__file__))
_RTMPOSE_URL = ("https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/"
"onnx_sdk/rtmpose-m_simcc-hand5_pt-aic-coco_210e-256x256-"
"74fb594_20230320.zip")
_DEFAULT_TASK = os.path.join(_PACKAGE_ROOT, "hand_landmarker.task")
_MP_PAD = 1.25
_MP_MIN_CONF = 0.3
class RtmposePoseBackend:
"""RTMPose hand5 SIMCC backend with the canonical call contract."""
name = "rtmpose"
def __init__(self, device: str = "cpu"):
if device == "auto":
# auto = 有 CUDA provider 用 cuda,否则 cpu(rtmlib 不识别 auto)
try:
import onnxruntime as ort
device = ("cuda" if "CUDAExecutionProvider"
in ort.get_available_providers() else "cpu")
except Exception:
device = "cpu"
try:
from rtmlib import RTMPose
self._pose = RTMPose(_RTMPOSE_URL, model_input_size=(256, 256),
backend="onnxruntime", device=device)
except Exception as exc:
print(f"RTMPose {device} 初始化失败（{exc}），回退 CPU")
from rtmlib import RTMPose
self._pose = RTMPose(_RTMPOSE_URL, model_input_size=(256, 256),
backend="onnxruntime", device="cpu")
device = "cpu"
self.device = device
def __call__(self, frame_bgr, bboxes=None):
return self._pose(frame_bgr, bboxes=bboxes)
def close(self):
self._pose = None
class MediaPipePoseBackend:
"""MediaPipe Tasks hand landmarker with bbox-to-hand association."""
name = "mediapipe"
def __init__(self, model_path: Optional[str] = None,
device: str = "cpu"):
import mediapipe as mp
from mediapipe.tasks import python as mp_python
from mediapipe.tasks.python import vision
self._mp = mp
self._vision = vision
self.model_path = model_path or _DEFAULT_TASK
if not os.path.isfile(self.model_path):
raise FileNotFoundError(self.model_path)
def make(delegate):
options = vision.HandLandmarkerOptions(
base_options=mp_python.BaseOptions(
model_asset_path=self.model_path, delegate=delegate),
running_mode=vision.RunningMode.IMAGE,
num_hands=2,
min_hand_detection_confidence=_MP_MIN_CONF,
min_hand_presence_confidence=_MP_MIN_CONF,
min_tracking_confidence=_MP_MIN_CONF,
)
return vision.HandLandmarker.create_from_options(options)
self._lm = make(mp_python.BaseOptions.Delegate.CPU)
if device in ("auto", "cuda"):
try:
candidate = make(mp_python.BaseOptions.Delegate.GPU)
_smoke(candidate, mp)
self._lm.close()
self._lm = candidate
device = "cuda"
except Exception as exc:
print(f"HandLandmarker GPU 初始化失败（{exc}），回退 CPU")
device = "cpu"
self.device = device
def __call__(self, frame_bgr, bboxes=None):
h, w = frame_bgr.shape[:2]
if not bboxes:
bboxes = [[0.0, 0.0, float(w), float(h)]]
rgb = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2RGB)
result = self._lm.detect(self._mp.Image(
image_format=self._mp.ImageFormat.SRGB, data=rgb))
kpts = np.zeros((len(bboxes), 21, 2), np.float32)
scores = np.zeros((len(bboxes), 21), np.float32)
if not result.hand_landmarks:
return kpts, scores
detected = []
for landmarks in result.hand_landmarks:
pts = np.array([[float(landmarks[j].x) * w,
float(landmarks[j].y) * h]
for j in range(21)], np.float32)
vis = np.array([
float(landmarks[j].visibility)
if landmarks[j].visibility is not None
else float(landmarks[j].presence)
if landmarks[j].presence is not None else 1.0
for j in range(21)
], np.float32)
detected.append((pts, vis))
used = [False] * len(detected)
for i, box in enumerate(bboxes):
x1, y1, x2, y2 = [float(v) for v in box]
bw, bh = x2 - x1, y2 - y1
pad = max(bw, bh) * (_MP_PAD - 1.0) / 2.0
best_j, best_d = -1, float("inf")
for j, (pts, _vis) in enumerate(detected):
if used[j]:
continue
cx, cy = float(pts[:, 0].mean()), float(pts[:, 1].mean())
if (cx < x1 - pad or cx > x2 + pad
or cy < y1 - pad or cy > y2 + pad):
continue
d = ((cx - (x1 + x2) / 2.0) ** 2
+ (cy - (y1 + y2) / 2.0) ** 2)
if d < best_d:
best_j, best_d = j, d
if best_j >= 0:
used[best_j] = True
kpts[i] = detected[best_j][0]
scores[i] = detected[best_j][1]
return kpts, scores
def close(self):
if self._lm is not None:
try:
self._lm.close()
except Exception:
pass
self._lm = None
def _smoke(landmarker, mp) -> None:
dummy = np.zeros((64, 64, 3), np.uint8)
landmarker.detect(mp.Image(image_format=mp.ImageFormat.SRGB,
data=dummy))
|