[Paper]
[arXiv]
[Project Page]
Localization with Sampling-Argmax
Jiefeng Li, Tong Chen, Ruiqi Shi, Yujing Lou, Yong-Lu Li, Cewu Lu
NeurIPS 2021
Differentiable Sampling
- Python 3.6+
- PyTorch >= 1.2
- torchvision >= 0.3.0
- Install PyTorch
conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0- Install
sampling_argmax
python setup.py developPlease download data from MSCOCO, Human3.6M and MTFL. Download and extract them under ./data, and make them look like this:
|-- exp
|-- sampling_argmax
|-- configs
|-- data
`-- |-- coco
|-- annotations
|-- train2017
`-- val2017
|-- h36m
|-- annotations
`-- images
|-- mtfl
`-- |-- AFLW
|---net_7876
|---lfw_5590
|-- training.json
`-- testing.json
# COCO Keypoint
./scripts/train_pose.sh configs/coco/256x192_res50_lr1e-3_1x-simple-integral.yaml coco_samp
# Human3.6M
./scripts/train_pose.sh configs/h36m/256x192_adam_lr1e-3-simple_3d_base_1x_h36mmpii.yaml h36m_samp
# MTFL
./scripts/train_mtfl.sh configs/mtfl/256x192_res50_lr1e-3_1x-mtfl-simple-integral.yaml mtfl_samp# COCO Keypoint
./scripts/validate_pose.sh configs/coco/256x192_res50_lr1e-3_1x-simple-integral.yaml ${CKPT}
# Human3.6M
./scripts/validate_pose.sh configs/h36m/256x192_adam_lr1e-3-simple_3d_base_1x_h36mmpii.yaml ${CKPT}
# MTFL
./scripts/validate_mtfl.sh configs/mtfl/256x192_res50_lr1e-3_1x-mtfl-simple-integral.yaml ${CKPT}Results on COCO validation set:
| Method | AP @0.5:0.95 | AP @0.5 | AP @0.75 |
|---|---|---|---|
| Samp. Uni. | 68.2 | 87.2 | 75.0 |
| Samp. Tri. | 69.8 | 87.9 | 76.2 |
| Samp. Gau. | 68.3 | 87.3 | 75.2 |
Results on S9 and S11:
| Method | MPJPE | PA-MPJPE |
|---|---|---|
| Samp. Uni. | 49.6 | 39.1 |
| Samp. Tri. | 49.5 | 39.1 |
| Samp. Gau. | 50.9 | 39.0 |
Results on MTFL:
| Method | Abs | Rel |
|---|---|---|
| Samp. Uni. | 3.00 | 6.86 |
| Samp. Tri. | 2.98 | 6.82 |
| Samp. Gau. | 2.94 | 6.96 |
If you find our code or paper useful, please consider citing
@inproceedings{li2021localization,
title={Localization with Sampling-Argmax},
author={Li, Jiefeng and Chen, Tong and Shi, Ruiqi and Lou, Yujing and Li, Yong-Lu and Lu, Cewu},
booktitle={Advances in Neural Information Processing Systems (NeurIPS)},
year={2021}
}