0602 MMPose 代码教学

王鼎康�lincoln 2023-06-05 01:52:55

RTMPose关键点检测全流程

保姆级代码教程： github.com/TommyZihao/MMPose_Tutorials

使用OpenMMLab开源目标检测算法库MMDetection，分别训练Faster R CNN和RTMDet的三角板目标检测算法。在测试集上评估精度，预测图像、视频、摄像头实时画面。目标检测预测框，用于后续Top-Down关键点检测算法。

流程

下载数据 -> 训练模型 -> 验证模型

安装

需要同时安装MMPose 和MMdetection。

预训练验证

python demo/topdown_demo_with_mmdet.py \
        demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py \
        https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
        projects/rtmpose/rtmpose/body_2d_keypoint/rtmpose-s_8xb256-420e_coco-256x192.py \
        https://download.openmmlab.com/mmpose/v1/projects/rtmpose/rtmpose-s_simcc-aic-coco_pt-aic-coco_420e-256x192-fcb2599b_20230126.pth \
        --input data/test/multi-person.jpeg \
        --output-root outputs/B1_RTM_1 \
        --device cuda:0 \
        --bbox-thr 0.5 \
        --kpt-thr 0.5 \
        --nms-thr 0.3 \
        --radius 8 \
        --thickness 4 \
        --draw-bbox \
        --draw-heatmap \
        --show-kpt-idx

python demo/topdown_demo_with_mmdet.py \
        demo/mmdetection_cfg/faster_rcnn_r50_fpn_coco.py \
        https://download.openmmlab.com/mmdetection/v2.0/faster_rcnn/faster_rcnn_r50_fpn_1x_coco/faster_rcnn_r50_fpn_1x_coco_20200130-047c8118.pth \
        configs/body_2d_keypoint/topdown_heatmap/coco/td-hm_hrnet-w32_8xb64-210e_coco-256x192.py \
        https://download.openmmlab.com/mmpose/top_down/hrnet/hrnet_w32_coco_256x192-c78dce93_20200708.pth \
        --input data/test/mother_wx.mp4 \
        --output-root outputs/B1_HRNet_2 \
        --device cuda:0 \
        --bbox-thr 0.5 \
        --kpt-thr 0.2 \
        --nms-thr 0.3 \
        --radius 5 \
        --thickness 2 \
        --draw-bbox \
        --draw-heatmap \
        --show-kpt-idx

获取数据

获取配置文件：

Faster-RCNN

https://zihao-openmmlab.obs.myhuaweicloud.com/20220610-mmpose/triangle_dataset/faster_r_cnn_triangle.py

RTMDet-tiny

https://zihao-openmmlab.obs.myhuaweicloud.com/20220610-mmpose/triangle_dataset/rtmdet_tiny_triangle.py

获取COCO标注的数据：

Triangle_215_Keypoint_coco$ tree -L 1
.
├── images
├── train_coco.json
└── val_coco.json

MMPose

https://zihao-openmmlab.obs.cn-east-3.myhuaweicloud.com/20220610-mmpose/triangle_dataset/rtmpose-s_triangle_8xb256-420e_coco-256x192.py

训练

MMdetection 训练

python tools/train.py data/rtmdet_tiny_triangle.py

训练日志文件在

# 日志文件路径
log_path = 'work_dirs/faster_r_cnn_triangle/20230603_120659/vis_data/scalars.json'

模型精简转换

# RTMDet-tiny
!python tools/model_converters/publish_model.py \
        work_dirs/rtmdet_tiny_triangle/epoch_200.pth \
        checkpoint/rtmdet_tiny_triangle_epoch_200_202305120847.pth

训练RTMPose

python tools/train.py data/rtmpose-s_triangle_8xb256-420e_coco-256x192.py

测试

检测（python）

# RTMDet 三角板检测
detector = init_detector(
    'data/rtmdet_tiny_triangle.py',
    'checkpoint/rtmdet_tiny_triangle_epoch_200_202305120847-3cd02a8f.pth',
    device=device
)

pose_estimator = init_pose_estimator(
    'data/rtmpose-s_triangle_8xb256-420e_coco-256x192.py',
    'checkpoint/rtmpose_s_triangle_300-34bfaeb2_20230512.pth',
    device=device,
    cfg_options={'model': {'test_cfg': {'output_heatmaps': True}}}
)

init_default_scope(detector.cfg.get('default_scope', 'mmdet'))

# 获取目标检测预测结果
detect_result = inference_detector(detector, img_path)

# 置信度
detect_result.pred_instances.scores

# 框坐标：左上角X坐标、左上角Y坐标、右下角X坐标、右下角Y坐标
detect_result.pred_instances.bboxes

pred_instance = detect_result.pred_instances.cpu().numpy()
bboxes = np.concatenate((pred_instance.bboxes, pred_instance.scores[:, None]), axis=1)
bboxes = bboxes[np.logical_and(pred_instance.labels == 0, pred_instance.scores > CONF_THRES)]
bboxes = bboxes[nms(bboxes, 0.3)][:, :4].astype('int')

可视化

# 半径
pose_estimator.cfg.visualizer.radius = 50
# 线宽
pose_estimator.cfg.visualizer.line_width = 20
visualizer = VISUALIZERS.build(pose_estimator.cfg.visualizer)
# 元数据
visualizer.set_dataset_meta(pose_estimator.dataset_meta)

img = mmcv.imread(img_path)
img = mmcv.imconvert(img, 'bgr', 'rgb')

img_output = visualizer.add_datasample(
            'result',
            img,
            data_sample=data_samples,
            draw_gt=False,
            draw_heatmap=True,
            draw_bbox=True,
            show=False,
            show_kpt_idx=True,
            wait_time=0,
            out_file='outputs/G3_visualizer.jpg',
            kpt_thr=0.3
)

plt.figure(figsize=(10, 10))
plt.imshow(img_output)
plt.show()

...全文