简体中文 | English
- Introduction
- Model Recommendation
- Model Zoo
- Getting Start
- Train with custom data
- BenchMark
The keypoint detection part in PaddleDetection follows the state-of-the-art algorithm closely, including Top-Down and Bottom-Up methods, which can satisfy the different needs of users. Top-Down detects the object first and then detects the specific keypoint. Top-Down models will be more accurate, but slower as the number of objects increases. Differently, Bottom-Up detects the point first and then group or connect those points to form several instances of human pose. The speed of Bottom-Up is fixed, it won't slow down as the number of objects increases, but it will be less accurate.
At the same time, PaddleDetection provides a self-developed real-time keypoint detection model PP-TinyPose optimized for mobile devices.
Detection Model | Keypoint Model | Input Size | Accuracy of COCO | Average Inference Time (FP16) | Params (M) | Flops (G) | Model Weight | Paddle-Lite Inference Model(FP16) |
---|---|---|---|---|---|---|---|---|
PicoDet-S-Pedestrian | PP-TinyPose | Detection:192x192 Keypoint:128x96 |
Detection mAP:29.0 Keypoint AP:58.1 |
Detection:2.37ms Keypoint:3.27ms |
Detection:1.18 Keypoint:1.36 |
Detection:0.35 Keypoint:0.08 |
Detection Keypoint |
Detection Keypoint |
PicoDet-S-Pedestrian | PP-TinyPose | Detection:320x320 Keypoint:256x192 |
Detection mAP:38.5 Keypoint AP:68.8 |
Detection:6.30ms Keypoint:8.33ms |
Detection:1.18 Keypoint:1.36 |
Detection:0.97 Keypoint:0.32 |
Detection Keypoint |
Detection Keypoint |
*Specific documents of PP-TinyPose, please refer to Document。
Detection Model | Keypoint Model | Input Size | Accuracy of COCO | Params (M) | Flops (G) | Model Weight |
---|---|---|---|---|---|---|
PP-YOLOv2 | HRNet-w32 | Detection:640x640 Keypoint:384x288 |
Detection mAP:49.5 Keypoint AP:77.8 |
Detection:54.6 Keypoint:28.6 |
Detection:115.8 Keypoint:17.3 |
Detection Keypoint |
PP-YOLOv2 | HRNet-w32 | Detection:640x640 Keypoint:256x192 |
Detection mAP:49.5 Keypoint AP:76.9 |
Detection:54.6 Keypoint:28.6 |
Detection:115.8 Keypoint:7.68 |
Detection Keypoint |
COCO Dataset
Note:1.The AP results of Top-Down models are based on bounding boxes in GroundTruth. 2.Vitpose training uses MAE as the pre-training model
MPII Dataset
Model | Input Size | PCKh(Mean) | PCKh(Mean@0.1) | Model Download | Config File |
---|---|---|---|---|---|
HRNet-w32 | 256x256 | 90.6 | 38.5 | hrnet_w32_256x256_mpii.pdparams | config |
Model for Scenes
Model | Strategy | Input Size | Precision | Inference Speed | Model Weights | Model Inference and Deployment | description |
---|---|---|---|---|---|---|---|
HRNet-w32 + DarkPose | Top-Down | 256x192 | AP: 87.1 (on internal dataset) | 2.9ms per person | Link | Link | Especially optimized for fall scenarios, the model is applied to PP-Human |
We also release PP-TinyPose, a real-time keypoint detection model optimized for mobile devices. Welcome to experience.
Please refer to PaddleDetection Installation Guide to install PaddlePaddle and PaddleDetection correctly.
Currently, KeyPoint Detection Models support COCO and MPII. Please refer to Keypoint Dataset Preparation to prepare dataset.
About the description for config files, please refer to Keypoint Config Guild.
- Note that, when testing by detected bounding boxes in Top-Down method, We should get
bbox.json
by a detection model. You can download the detected results for COCO val2017 (Detector having human AP of 56.4 on COCO val2017 dataset) directly, put it at the root path (PaddleDetection/
), and setuse_gt_bbox: False
in config file.
#COCO DataSet
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/keypoint/higherhrnet/higherhrnet_hrnet_w32_512.yml
#MPII DataSet
CUDA_VISIBLE_DEVICES=0 python3 tools/train.py -c configs/keypoint/hrnet/hrnet_w32_256x256_mpii.yml
#COCO DataSet
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m paddle.distributed.launch tools/train.py -c configs/keypoint/higherhrnet/higherhrnet_hrnet_w32_512.yml
#MPII DataSet
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 -m paddle.distributed.launch tools/train.py -c configs/keypoint/hrnet/hrnet_w32_256x256_mpii.yml
#COCO DataSet
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/keypoint/higherhrnet/higherhrnet_hrnet_w32_512.yml
#MPII DataSet
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/keypoint/hrnet/hrnet_w32_256x256_mpii.yml
#If you only need the prediction result, you can set --save_prediction_only. Then the result will be saved at output/keypoints_results.json by default.
CUDA_VISIBLE_DEVICES=0 python3 tools/eval.py -c configs/keypoint/higherhrnet/higherhrnet_hrnet_w32_512.yml --save_prediction_only
Note:Top-down models only support inference for a cropped image with single person. If you want to do inference on image with several people, please see "joint inference by detection and keypoint". Or you can choose a Bottom-up model.
CUDA_VISIBLE_DEVICES=0 python3 tools/infer.py -c configs/keypoint/higherhrnet/higherhrnet_hrnet_w32_512.yml -o weights=./output/higherhrnet_hrnet_w32_512/model_final.pdparams --infer_dir=../images/ --draw_threshold=0.5 --save_txt=True
#Export Detection Model
python tools/export_model.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams
#Export Keypoint Model
python tools/export_model.py -c configs/keypoint/hrnet/hrnet_w32_256x192.yml -o weights=https://paddledet.bj.bcebos.com/models/keypoint/hrnet_w32_256x192.pdparams
#Deployment for detector and keypoint, which is only for Top-Down models
python deploy/python/det_keypoint_unite_infer.py --det_model_dir=output_inference/ppyolo_r50vd_dcn_2x_coco/ --keypoint_model_dir=output_inference/hrnet_w32_384x288/ --video_file=../video/xxx.mp4 --device=gpu
#Export model
python tools/export_model.py -c configs/keypoint/higherhrnet/higherhrnet_hrnet_w32_512.yml -o weights=output/higherhrnet_hrnet_w32_512/model_final.pdparams
#Keypoint independent deployment, which is only for bottom-up models
python deploy/python/keypoint_infer.py --model_dir=output_inference/higherhrnet_hrnet_w32_512/ --image_file=./demo/000000014439_640x640.jpg --device=gpu --threshold=0.5
#export FairMOT model
python tools/export_model.py -c configs/mot/fairmot/fairmot_dla34_30e_1088x608.yml -o weights=https://paddledet.bj.bcebos.com/models/mot/fairmot_dla34_30e_1088x608.pdparams
#joint inference with Multi-Object Tracking model FairMOT
python deploy/python/mot_keypoint_unite_infer.py --mot_model_dir=output_inference/fairmot_dla34_30e_1088x608/ --keypoint_model_dir=output_inference/higherhrnet_hrnet_w32_512/ --video_file={your video name}.mp4 --device=GPU
Note: To export MOT model, please refer to Here.
We provide standalone deploy of PaddleInference(Server-GPU)、PaddleLite(mobile、ARM)、Third-Engine(MNN、OpenVino), which is independent of training codes。For detail, please click Deploy-docs。
We take an example of tinypose_256x192 to show how to train with custom data.
1、For configs tinypose_256x192.yml
you may need to modify these for your job:
num_joints: &num_joints 17 #the number of joints in your job
train_height: &train_height 256 #the height of model input
train_width: &train_width 192 #the width of model input
hmsize: &hmsize [48, 64] #the shape of model output,usually 1/4 of [w,h]
flip_perm: &flip_perm [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]] #the correspondence between left and right keypoint id,used for flip transform。You can add an line(by "flip: False") behind of flip_pairs in RandomFlipHalfBodyTransform of TrainReader if you don't need it
num_joints_half_body: 8 #The joint numbers of half body, used for half_body transform
prob_half_body: 0.3 #The probability of half_body transform, set to 0 if you don't need it
upper_body_ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] #The joint ids of half(upper) body, used to get the upper joints in half_body transform
For more configs, please refer to KeyPointConfigGuide。
- In keypoint_utils.py, please set: "sigmas = np.array([.26, .25, .25, .35, .35, .79, .79, .72, .72, .62, .62, 1.07, 1.07,.87, .87, .89, .89]) / 10.0", the value indicate the variance of a joint locations,normally 0.25-0.5 means the location is highly accuracy,for example: eyes。0.5-1.0 means the location is not sure so much,for example: shoulder。0.75 is recommand if you not sure。
- In visualizer.py, please set "EDGES" in draw_pose function,this indicate the line to show between joints for visualization。
- In pycocotools you installed, please set "sigmas",it is the same as that in keypoint_utils.py, but used for coco evaluation。
- The data should has the same format as Coco data, and the keypoints(Nx3) and bbox(N) should be annotated.
- please set "area">0 in annotations files otherwise it will be skiped while training. Moreover, due to the evaluation mechanism of COCO, the data with small area may also be filtered out during evaluation. We recommend to set
area = bbox_w * bbox_h
when customizing your dataset.
We provide benchmarks in different runtime environments for your reference when choosing models. See Keypoint Inference Benchmark for details.
@inproceedings{cheng2020bottom,
title={HigherHRNet: Scale-Aware Representation Learning for Bottom-Up Human Pose Estimation},
author={Bowen Cheng and Bin Xiao and Jingdong Wang and Honghui Shi and Thomas S. Huang and Lei Zhang},
booktitle={CVPR},
year={2020}
}
@inproceedings{SunXLW19,
title={Deep High-Resolution Representation Learning for Human Pose Estimation},
author={Ke Sun and Bin Xiao and Dong Liu and Jingdong Wang},
booktitle={CVPR},
year={2019}
}
@article{wang2019deep,
title={Deep High-Resolution Representation Learning for Visual Recognition},
author={Wang, Jingdong and Sun, Ke and Cheng, Tianheng and Jiang, Borui and Deng, Chaorui and Zhao, Yang and Liu, Dong and Mu, Yadong and Tan, Mingkui and Wang, Xinggang and Liu, Wenyu and Xiao, Bin},
journal={TPAMI},
year={2019}
}
@InProceedings{Zhang_2020_CVPR,
author = {Zhang, Feng and Zhu, Xiatian and Dai, Hanbin and Ye, Mao and Zhu, Ce},
title = {Distribution-Aware Coordinate Representation for Human Pose Estimation},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}
@inproceedings{Yulitehrnet21,
title={Lite-HRNet: A Lightweight High-Resolution Network},
author={Yu, Changqian and Xiao, Bin and Gao, Changxin and Yuan, Lu and Zhang, Lei and Sang, Nong and Wang, Jingdong},
booktitle={CVPR},
year={2021}
}
@inproceedings{
xu2022vitpose,
title={ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation},
author={Yufei Xu and Jing Zhang and Qiming Zhang and Dacheng Tao},
booktitle={Advances in Neural Information Processing Systems},
year={2022},
}