camenduru/vitpose

The official repo for [NeurIPS'22] "ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation" and [Arxiv'22] "ViTPose+: Vision Transformer Foundation Model for Generic Body Pose Estimation"

ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation

Results | Updates | Usage | Todo | Acknowledge

This branch contains the pytorch implementation of ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation and ViTPose+: Vision Transformer Foundation Model for Generic Body Pose Estimation. It obtains 81.1 AP on MS COCO Keypoint test-dev set.

Web Demo

• Integrated into Huggingface Spaces 🤗 using Gradio. Try out the Web Demo for video: and images

MAE Pre-trained model

• The small size MAE pre-trained model can be found in Onedrive.
• The base, large, and huge pre-trained models using MAE can be found in the MAE official repo.

Results from this repo on MS COCO val set (single-task training)

Using detection results from a detector that obtains 56 mAP on person. The configs here are for both training and test.

With classic decoder

Model	Pretrain	Resolution	AP	AR	config	log	weight
ViTPose-S	MAE	256x192	73.8	79.2	config	log	Onedrive
ViTPose-B	MAE	256x192	75.8	81.1	config	log	Onedrive
ViTPose-L	MAE	256x192	78.3	83.5	config	log	Onedrive
ViTPose-H	MAE	256x192	79.1	84.1	config	log	Onedrive

With simple decoder

Model	Pretrain	Resolution	AP	AR	config	log	weight
ViTPose-S	MAE	256x192	73.5	78.9	config	log	Onedrive
ViTPose-B	MAE	256x192	75.5	80.9	config	log	Onedrive
ViTPose-L	MAE	256x192	78.2	83.4	config	log	Onedrive
ViTPose-H	MAE	256x192	78.9	84.0	config	log	Onedrive

Results with multi-task training

Note * There may exist duplicate images in the crowdpose training set and the validation images in other datasets, as discussed in issue #24. Please be careful when using these models for evaluation. We provide the results without the crowpose dataset for reference.

Human datasets (MS COCO, AIC, MPII, CrowdPose)

Results on MS COCO val set

Using detection results from a detector that obtains 56 mAP on person. Note the configs here are only for evaluation.

Model	Dataset	Resolution	AP	AR	config	weight
ViTPose-B	COCO+AIC+MPII	256x192	77.1	82.2	config	Onedrive
ViTPose-L	COCO+AIC+MPII	256x192	78.7	83.8	config	Onedrive
ViTPose-H	COCO+AIC+MPII	256x192	79.5	84.5	config	Onedrive
ViTPose-G	COCO+AIC+MPII	576x432	81.0	85.6
ViTPose-B*	COCO+AIC+MPII+CrowdPose	256x192	77.5	82.6	config	Onedrive
ViTPose-L*	COCO+AIC+MPII+CrowdPose	256x192	79.1	84.1	config	Onedrive
ViTPose-H*	COCO+AIC+MPII+CrowdPose	256x192	79.8	84.8	config	Onedrive
ViTPose+-S	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	75.8	82.6	config	log | Onedrive
ViTPose+-B	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	77.0	82.6	config	log | Onedrive
ViTPose+-L	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	78.6	84.1	config	log | Onedrive
ViTPose+-H	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	79.4	84.8	config	log | Onedrive

Results on OCHuman test set

Using groundtruth bounding boxes. Note the configs here are only for evaluation.

Model	Dataset	Resolution	AP	AR	config	weight
ViTPose-B	COCO+AIC+MPII	256x192	88.0	89.6	config	Onedrive
ViTPose-L	COCO+AIC+MPII	256x192	90.9	92.2	config	Onedrive
ViTPose-H	COCO+AIC+MPII	256x192	90.9	92.3	config	Onedrive
ViTPose-G	COCO+AIC+MPII	576x432	93.3	94.3
ViTPose-B*	COCO+AIC+MPII+CrowdPose	256x192	88.2	90.0	config	Onedrive
ViTPose-L*	COCO+AIC+MPII+CrowdPose	256x192	91.5	92.8	config	Onedrive
ViTPose-H*	COCO+AIC+MPII+CrowdPose	256x192	91.6	92.8	config	Onedrive
ViTPose+-S	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	78.4	80.6	config	log | Onedrive
ViTPose+-B	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	82.6	84.8	config	log | Onedrive
ViTPose+-L	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	85.7	87.5	config	log | Onedrive
ViTPose+-H	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	85.7	87.4	config	log | Onedrive

Results on MPII val set

Using groundtruth bounding boxes. Note the configs here are only for evaluation. The metric is PCKh.

Model	Dataset	Resolution	Mean	config	weight
ViTPose-B	COCO+AIC+MPII	256x192	93.3	config	Onedrive
ViTPose-L	COCO+AIC+MPII	256x192	94.0	config	Onedrive
ViTPose-H	COCO+AIC+MPII	256x192	94.1	config	Onedrive
ViTPose-G	COCO+AIC+MPII	576x432	94.3
ViTPose-B*	COCO+AIC+MPII+CrowdPose	256x192	93.4	config	Onedrive
ViTPose-L*	COCO+AIC+MPII+CrowdPose	256x192	93.9	config	Onedrive
ViTPose-H*	COCO+AIC+MPII+CrowdPose	256x192	94.1	config	Onedrive
ViTPose+-S	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	92.7	config	log | Onedrive
ViTPose+-B	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	92.8	config	log | Onedrive
ViTPose+-L	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	94.0	config	log | Onedrive
ViTPose+-H	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	94.2	config	log | Onedrive

Results on AI Challenger test set

Using groundtruth bounding boxes. Note the configs here are only for evaluation.

Model	Dataset	Resolution	AP	AR	config	weight
ViTPose-B	COCO+AIC+MPII	256x192	32.0	36.3	config	Onedrive
ViTPose-L	COCO+AIC+MPII	256x192	34.5	39.0	config	Onedrive
ViTPose-H	COCO+AIC+MPII	256x192	35.4	39.9	config	Onedrive
ViTPose-G	COCO+AIC+MPII	576x432	43.2	47.1
ViTPose-B*	COCO+AIC+MPII+CrowdPose	256x192	31.9	36.3	config	Onedrive
ViTPose-L*	COCO+AIC+MPII+CrowdPose	256x192	34.6	39.0	config	Onedrive
ViTPose-H*	COCO+AIC+MPII+CrowdPose	256x192	35.3	39.8	config	Onedrive
ViTPose+-S	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	29.7	34.3	config	log | Onedrive
ViTPose+-B	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	31.8	36.3	config	log | Onedrive
ViTPose+-L	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	34.3	38.9	config	log | Onedrive
ViTPose+-H	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	34.8	39.1	config	log | Onedrive

Results on CrowdPose test set

Using YOLOv3 human detector. Note the configs here are only for evaluation.

Model	Dataset	Resolution	AP	AP(H)	config	weight
ViTPose-B*	COCO+AIC+MPII+CrowdPose	256x192	74.7	63.3	config	Onedrive
ViTPose-L*	COCO+AIC+MPII+CrowdPose	256x192	76.6	65.9	config	Onedrive
ViTPose-H*	COCO+AIC+MPII+CrowdPose	256x192	76.3	65.6	config	Onedrive

Animal datasets (AP10K, APT36K)

Results on AP-10K test set

Model	Dataset	Resolution	AP	config	weight
ViTPose+-S	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	71.4	config	log | Onedrive
ViTPose+-B	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	74.5	config	log | Onedrive
ViTPose+-L	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	80.4	config	log | Onedrive
ViTPose+-H	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	82.4	config	log | Onedrive

Results on APT-36K val set

Model	Dataset	Resolution	AP	config	weight
ViTPose+-S	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	74.2	config	log | Onedrive
ViTPose+-B	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	75.9	config	log | Onedrive
ViTPose+-L	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	80.8	config	log | Onedrive
ViTPose+-H	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	82.3	config	log | Onedrive

WholeBody dataset

Model	Dataset	Resolution	AP	config	weight
ViTPose+-S	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	54.4	config	log | Onedrive
ViTPose+-B	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	57.4	config	log | Onedrive
ViTPose+-L	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	60.6	config	log | Onedrive
ViTPose+-H	COCO+AIC+MPII+AP10K+APT36K+WholeBody	256x192	61.2	config	log | Onedrive

Transfer results on the hand dataset (InterHand2.6M)

Model	Dataset	Resolution	AUC	config	weight
ViTPose+-S	COCO+AIC+MPII+WholeBody	256x192	86.5	config	Coming Soon
ViTPose+-B	COCO+AIC+MPII+WholeBody	256x192	87.0	config	Coming Soon
ViTPose+-L	COCO+AIC+MPII+WholeBody	256x192	87.5	config	Coming Soon
ViTPose+-H	COCO+AIC+MPII+WholeBody	256x192	87.6	config	Coming Soon

Updates

[2023-01-10] Update ViTPose+! It uses MoE strategies to jointly deal with human, animal, and wholebody pose estimation tasks.

[2022-05-24] Upload the single-task training code, single-task pre-trained models, and multi-task pretrained models.

[2022-05-06] Upload the logs for the base, large, and huge models!

[2022-04-27] Our ViTPose with ViTAE-G obtains 81.1 AP on COCO test-dev set!

Applications of ViTAE Transformer include: image classification | object detection | semantic segmentation | animal pose segmentation | remote sensing | matting | VSA | ViTDet

Usage

We use PyTorch 1.9.0 or NGC docker 21.06, and mmcv 1.3.9 for the experiments.

git clone https://github.com/open-mmlab/mmcv.git
cd mmcv
git checkout v1.3.9
MMCV_WITH_OPS=1 pip install -e .
cd ..
git clone https://github.com/ViTAE-Transformer/ViTPose.git
cd ViTPose
pip install -v -e .

After install the two repos, install timm and einops, i.e.,

pip install timm==0.4.9 einops

After downloading the pretrained models, please conduct the experiments by running

# for single machine
bash tools/dist_train.sh <Config PATH> <NUM GPUs> --cfg-options model.pretrained=<Pretrained PATH> --seed 0

# for multiple machines
python -m torch.distributed.launch --nnodes <Num Machines> --node_rank <Rank of Machine> --nproc_per_node <GPUs Per Machine> --master_addr <Master Addr> --master_port <Master Port> tools/train.py <Config PATH> --cfg-options model.pretrained=<Pretrained PATH> --launcher pytorch --seed 0

To test the pretrained models performance, please run

bash tools/dist_test.sh <Config PATH> <Checkpoint PATH> <NUM GPUs>

For ViTPose+ pre-trained models, please first re-organize the pre-trained weights using

python tools/model_split.py --source <Pretrained PATH>

Todo

This repo current contains modifications including:

• Upload configs and pretrained models

• More models with SOTA results

• Upload multi-task training config

Acknowledge

We acknowledge the excellent implementation from mmpose and MAE.

Citing ViTPose

For ViTPose

@inproceedings{
  xu2022vitpose,
  title={Vi{TP}ose: 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},
}

For ViTPose+

@article{xu2022vitpose+,
  title={ViTPose+: Vision Transformer Foundation Model for Generic Body Pose Estimation},
  author={Xu, Yufei and Zhang, Jing and Zhang, Qiming and Tao, Dacheng},
  journal={arXiv preprint arXiv:2212.04246},
  year={2022}
}

For ViTAE and ViTAEv2, please refer to:

@article{xu2021vitae,
  title={Vitae: Vision transformer advanced by exploring intrinsic inductive bias},
  author={Xu, Yufei and Zhang, Qiming and Zhang, Jing and Tao, Dacheng},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}

@article{zhang2022vitaev2,
  title={ViTAEv2: Vision Transformer Advanced by Exploring Inductive Bias for Image Recognition and Beyond},
  author={Zhang, Qiming and Xu, Yufei and Zhang, Jing and Tao, Dacheng},
  journal={arXiv preprint arXiv:2202.10108},
  year={2022}
}