HRNet-W48 Whole-body Keypoint Detection

# 基础依赖
pip install torch torchvision opencv-python numpy matplotlib

# 昇腾 NPU 依赖（可选）
pip install torch-npu decorator psutil

# 通过 ModelScope
pip install modelscope
modelscope download --model iic/cv_hrnetw48_human-wholebody-keypoint_image --local_dir ./weights

# 或手动下载至 weights/ 目录
# weights/yolox_x_large_human_wbdl.pth
# weights/hrnet_w48_wbdl.pth

cd hrnetw48-wholebody-keypoint

# 单张图片推理
python inference.py --image path/to/image.jpg

# 自动精度对比（CPU vs NPU）
python inference.py --compare

# 查看帮助
python inference.py --help

from inference import (
    build_detection_model, build_keypoint_model,
    preprocess_detection, postprocess_detection,
    preprocess_keypoint, decode_heatmaps,
    map_keypoints_to_original
)
import cv2
import torch

# 构建模型
det_model = build_detection_model('cpu')
kpt_model = build_keypoint_model('cpu')

# 读取图像
img = cv2.imread('image.jpg')

# 人体检测
img_tensor, scale, oh, ow, nh, nw = preprocess_detection(img)
with torch.no_grad():
    det_out = det_model(img_tensor)
boxes, scores, labels = postprocess_detection(det_out, scale, oh, ow, nh, nw)

# 关键点检测
img_rgb = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
for box in boxes:
    kpt_input, crop_box, crop_wh = preprocess_keypoint(img_rgb, box)
    with torch.no_grad():
        heatmap = kpt_model(kpt_input)
    kps = decode_heatmaps(heatmap)
    kps = map_keypoints_to_original(kps, crop_box, crop_wh, img.shape)
    print(f"133 keypoints: {kps.shape}")

[Step 1/6] 检查模型权重...                    ✓
[Step 2/6] 测试Python导入...                  ✓
[Step 3/6] 测试模型权重加载...                ✓
[Step 4/6] 测试前向推理...                    ✓  (187个检测框, 133个关键点)
[Step 5/6] CPU vs NPU 精度对比...             ✓
[Step 6/6] 性能基准测试...                     ✓

Device: npu:0
Image: 640x425
[Detection] 0 persons (431ms)              ← 密集街景大图，YOLOX 微调模型 obj 分支被抑制
[INFO] Using center crop fallback            ← 自动回退到中心裁剪
[Keypoint] 133 joints (47ms)                 ← HRNet-W48 全身关键点输出
Confidence: [0.668, 1.028]

  nose:       ( 367.5,  80.6) conf=1.004      leye:       ( 375.9,  76.2) conf=0.985
  reye:       ( 361.9,  76.2) conf=0.975      neck:       ( 406.7, 111.5) conf=0.884
  rshoulder:  ( 361.9, 124.8) conf=0.908      relbow:     ( 445.8, 168.9) conf=0.668
  rwrist:     ( 342.3, 160.1) conf=0.873      lshoulder:  ( 451.4, 166.7) conf=0.877
  lelbow:     ( 305.9, 177.8) conf=0.947      lwrist:     ( 431.8, 213.1) conf=0.878
  rhip:       ( 401.1, 215.3) conf=0.902      rknee:      ( 431.8, 290.4) conf=0.923
  rankle:     ( 373.1, 272.7) conf=0.955      lhip:       ( 473.8, 361.1) conf=0.849
  lknee:      ( 403.9, 341.2) conf=0.856      lankle:     ( 443.0, 376.5) conf=0.750

Visualization saved to /tmp/keypoint_result.jpg
[DONE]

hrnetw48-wholebody-keypoint/
├── inference.py        # 主推理脚本（模型定义 + 预处理 + 后处理 + 精度对比）
├── README.md           # 本文件
└── weights/            # 模型权重目录
    ├── yolox_x_large_human_wbdl.pth
    └── hrnet_w48_wbdl.pth