This repo aims to integral many fantastic works in human pose estimation with pytorch. This codes are heavily borrowed from HRNet and MSPNet, the code structure imitates the maskrcnn-benchmark, aims to fast, modular and distributed training with pytorch!
| model | pretrained | dataset | optimizer | iteration | eval AP | paper | Augment |
|---|---|---|---|---|---|---|---|
| MSPN | N | COCO | SGD | 96k | 72.5 | ||
| MSPN | N | COCO | Adam | 96k | 73.5 | ||
| MSPN | Y | COCO | Adam | 96k | 74.7 | ||
| MSPN | Y | COCO | Adam(paper) | 96k | 74.6 | 74.5 | |
| RES-50 | Y | COCO | Adam | 96k | 70.9 | ||
| RES-101 | Y | COCO | Adam | 96k | 70.1 | ||
| HR-W48 | Y | COCO | Adam | 96k | 74.2 | ||
| HR-W48 | Y | COCO | Adam | 96k | 72.1 | * | Y |
| EFFN-b4 | Y | COCO | Adam | 96k | 71.4 | ||
| EFFN-b4 | Y | COCO | Adam | 96k | 71.6 | * | Y |
| MSPN | Y | MPII | Adam(paper) | 28.8k | 90.17 | ||
| RES-50 | Y | MPII | Adam | 28.8k | 88.94 | ||
| HR-W48 | Y | MPII | Adam | 28.8k | 90.87 | ||
| EFFN-b4 | Y | MPII | Adam | 28.8k | 88.88 |
- coco dataset support
- add MSPN
- add resnet
- add hrnet
- add efficientnet
- mpii dataset support
Train the model by yourself is easily with this repo.
- Clone this repo
git clone https://github.com/tkianai/human-pose-estimation.pytorch- Prepare the data
# overall
datasets
├── coco
└── mpii
# coco
├── annotations
├── captions_train2017.json
├── captions_val2017.json
├── instances_train2017.json
├── instances_val2017.json
├── person_keypoints_train2017.json
├── person_keypoints_val2017_det.json.json
└── person_keypoints_val2017.json
├── train2017
├── val2017
# mpii
├── annotations
├── test.json
├── train.json
├── valid.json
└── valid.mat
└── images- Train
# single GPU training
python train.py --config-file configs/coco_mspn.yaml
# distributed training
python -m torch.distributed.launch --nproc_per_node=8 train.py --config-file configs/coco_mspn.yaml- Test
# single GPU testing
python test.py --config-file configs/coco_mspn.yaml --ckpt <model location>
# distributed testing
python -m torch.distributed.launch --nproc_per_node=8 test.py --config-file configs/coco_mspn.yaml --ckpt <model location>Try my best to keep update the new model structure, datasets and training strategies. Please star this repo~~~
New pull requests are strongly welcomed!
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent