NanoDet Face-Human-Hand Detection — 昇腾 NPU 部署

# 通过 ModelScope SDK 下载
pip install modelscope
python -c "
from modelscope import snapshot_download
snapshot_download('iic/cv_nanodet_face-human-hand-detection',
                  cache_dir='./models')
"

# 自动选择设备（NPU 优先）
python inference.py --image test.jpg

# 指定 CPU
python inference.py --image test.jpg --device cpu

# 指定 NPU
python inference.py --image test.jpg --device npu

# 自定义参数
python inference.py \
    --model-path models/iic/cv_nanodet_face-human-hand-detection/pytorch_model.bin \
    --image test.jpg \
    --device npu \
    --input-size 320 \
    --score-thresh 0.3 \
    --output-dir ./output

python inference.py --compare

python inference.py --benchmark --bench-iters 100

from inference import load_model, inference, NanoDetPreProcessor, NanoDetPostProcessor

# 加载模型
model = load_model('models/iic/cv_nanodet_face-human-hand-detection/pytorch_model.bin',
                   device='npu:0')

# 推理
labels, bboxes, scores = inference(model, 'test.jpg', device='npu:0',
                                    input_size=320, score_thresh=0.3)

# 结果解读
# labels:  [0, 1, 2] → [人体, 人脸, 人手]
# bboxes:  [[x1,y1,x2,y2], ...] 像素坐标
# scores:  [0.96, 0.89, 0.82] 置信度

$ python inference.py --image test_face_human_hand_detection.jpg --device npu

Loading model from: models/iic/cv_nanodet_face-human-hand-detection/pytorch_model.bin
Device: npu
  Loaded 645/645 parameters from checkpoint
Model loaded successfully!

Warming up...
Warm-up done.

Running inference on: test_face_human_hand_detection.jpg
Inference time: 162.72 ms

============================================================
Detection Results (score > 0.3):
============================================================
  Class                   Score BBox (x1,y1,x2,y2)
  ------------------------------------------------------------
  人体(body)               0.9692 [0.0, 0.1, 368.0, 640.0]
  人脸(face)               0.9165 [127.2, 82.9, 332.5, 366.2]
  人手(hand)               0.8349 [75.5, 286.1, 240.8, 511.5]

Done! Device=npu, Inference=162.72ms, Detections=3
Saved result to: output/result_npu.jpg

$ python inference.py --image test_face_human_hand_detection.jpg --device cpu

Loading model from: models/iic/cv_nanodet_face-human-hand-detection/pytorch_model.bin
Device: cpu
  Loaded 645/645 parameters from checkpoint

Running inference on: test_face_human_hand_detection.jpg
Inference time: 74.19 ms

============================================================
Detection Results (score > 0.3):
============================================================
  Class                   Score BBox (x1,y1,x2,y2)
  ------------------------------------------------------------
  人体(body)               0.9691 [0.0, 0.1, 368.0, 640.0]
  人脸(face)               0.9165 [127.2, 82.9, 332.5, 366.2]
  人手(hand)               0.8350 [75.5, 286.1, 240.8, 511.5]

Done! Device=cpu, Inference=74.19ms, Detections=3
Saved result to: output/result_cpu.jpg

$ python inference.py --compare

CPU vs NPU Precision Comparison (feature-level)
============================================================
  cls_score level 0: max_diff=0.013860, mean_diff=0.000810, cos_sim=1.000000
  cls_score level 1: max_diff=0.005308, mean_diff=0.000906, cos_sim=1.000000
  cls_score level 2: max_diff=0.003007, mean_diff=0.000804, cos_sim=1.000000
  cls_score level 3: max_diff=0.001945, mean_diff=0.000782, cos_sim=1.000000
  bbox_pred level 0: max_diff=0.020175, mean_diff=0.001067, cos_sim=0.999999
  bbox_pred level 1: max_diff=0.014530, mean_diff=0.001104, cos_sim=1.000000
  bbox_pred level 2: max_diff=0.008117, mean_diff=0.001090, cos_sim=1.000000
  bbox_pred level 3: max_diff=0.002050, mean_diff=0.000483, cos_sim=1.000000

  Matched detections: 3
  Score difference: mean=0.000054, max=0.000059

$ python inference.py --benchmark --bench-iters 30

============================================================
Performance Benchmark: device=cpu, input_size=320, iters=30
============================================================
  FPS: 21.0

============================================================
Performance Benchmark: device=npu:0, input_size=320, iters=30
============================================================
  FPS: 95.8

nanodet-face-human-hand/
├── inference.py          # 推理脚本（模型定义 + 推理 + 评估 + 基准测试）
├── README.md             # 部署文档
├── models/               # 模型权重
│   └── iic/cv_nanodet_face-human-hand-detection/
│       ├── pytorch_model.bin
│       └── test_face_human_hand_detection.jpg
└── output/               # 推理输出
    ├── result_cpu.jpg
    ├── result_npu.jpg
    ├── precision_comparison.json
    ├── precision_detail.json
    └── benchmark.json

Input (3×320×320, BGR)
  │
  ▼
ShuffleNetV2-1.0x Backbone
  ├─ stage2 → C2 (116, 80×80)
  ├─ stage3 → C3 (232, 40×40)
  └─ stage4 → C4 (464, 20×20)
        │
        ▼
GhostPAN Neck
  ├─ P4 (96, 20×20)  ← top-down from C4
  ├─ P3 (96, 40×40)  ← top-down from C3 + P4
  └─ P2 (96, 80×80)  ← top-down from C2 + P3
        │
        ▼
NanoDetPlus Head (GFL)
  ├─ cls_score: 4 levels × num_classes
  └─ bbox_pred:  4 levels × 4×(reg_max+1)
        │
        ▼
Post-processing
  ├─ Integral (softmax → weighted sum → distance)
  ├─ distance2bbox (anchor + distance → xyxy)
  ├─ Per-class score filtering (thresh=0.3)
  └─ NMS (IoU thresh=0.5)