First, follow the default instruction to install the project and datasets/README.md set up the datasets (e.g., MS-COCO).
For training on COCO, run:
python tools/train_net.py \
--num-gpus 8 \
--config-file configs/FCPose/R_50_3X.yaml \
--dist-url tcp://127.0.0.1:$(( RANDOM % 1000 + 50000 )) \
OUTPUT_DIR training_dir/R_50_3X
For evaluation on COCO, run:
python tools/train_net.py \
--num-gpus 8 \
--eval-only \
--config-file configs/FCPose/R_50_3X.yaml \
--dist-url tcp://127.0.0.1:$(( RANDOM % 1000 + 50000 )) \
OUTPUT_DIR training_dir/R_50_3X \
MODEL.WEIGHTS training_dir/R_50_3X/model_final.pth
Name | inf. time | box AP | mask AP | download |
---|---|---|---|---|
FCPose_R50_3x | 45ms | 57.9 | 65.2 | model |
FCPose_R101_3x | 58ms | 58.7 | 67.0 | model |
Disclaimer:
- Inference time is measured on 8 V100 GPUs.
- This is a reimplementation. Thus, the numbers are slightly different from our original paper.
- This is a alpha version. We will update our implement later, including adding real-time version FCPose and fixing the issue of the loss being nan. if you found you loss being nan when training, please try again.
Please consider citing our papers in your publications if the project helps your research. BibTeX reference is as follows.
@inproceedings{mao2021fcpose,
title={FCPose: Fully Convolutional Multi-Person Pose Estimation with Dynamic Instance-Aware Convolutions},
author={Mao, Weian and Tian, Zhi and Wang, Xinlong and Shen, Chunhua},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={9034--9043},
year={2021}
}