paod's Introduction

PAOD: Rethinking Prediction Alignment in One-stage Object Detection

The official implementation of the paper Rethinking Prediction Alignment in One-stage Object Detection.

News

2022.08.09: We release the code and models of PAOD.

Model Zoo

COCO

Model	Backbone	Lr Schd	mAP	AP50	AP75	Config	Model
PAOD	ResNeXt101	2x	48.8	67.3	53.3	Config	Google Drive
PAOD	ResNeXt101-DCN	2x	50.4	68.9	55.0	Config	Google Drive
PAOD	Res2Net-DCN	2x	51.1	69.6	55.8	Config	Google Drive

Pascal VOC

Model	Backbone	Lr Schd	mAP	AP50	AP75	Config	Model
PAOD	ResNet50	1x	65.0	85.6	71.2	Config	Google Drive

CrowdHuman

Detector	Backbone	Lr Schd	AP ↑	MR ↓	JI ↑	Config	Model
PAOD	ResNet50	1x	89.2	46.5	77.7	Config	Google Drive

Requirements

Please check installation for installation.

Training, Evaluation and Visualization

To train PAOD with 8 GPUs, run:

bash tools/dist_train.sh $CONFIG 8

or you can run the .sh file in script:

bash train_on_coco.sh
bash train_on_voc.sh
bash train_on_crow.sh

To evaluate PAOD with 8 GPU, run:

bash tools/dist_test.sh $YOUR_CONFIG $YOUR_CKPT 8 --eval=bbox

To visualize the predictions, run:

python tools/test.py $YOUR_CONFIG $YOUR_CKPT --eval=bbox --show

Citation

We appreciate it if you would please cite the following paper if you found the implementation useful for your work:

@article{Xiao2022RethinkingPA,
  title={Rethinking Prediction Alignment in One-stage Object Detection},
  author={Junrui Xiao and He Jiang and Zhikai Li and Qingyi Gu},
  journal={Neurocomputing},
  year={2022}
}

Acknowledgement

This project is mainly based on the following open-sourced projects: open-mmlab, and we thank DDOD for their code on CrowdHuman.

paod's People

Contributors

Stargazers

Watchers

paod's Issues

Some theoretical problems want to learn from you

Thank you for your excellent work about PAOD!!!
I have some question about theoretical problems want to learn from you.

compare with TOOD，you use DCN enhance with two concat features after decouple two task,for enhancing two task' consistence ,but TOOD do these things with share feature, What are the advantages of our approach？
How do you choose iou branch instead of GFL/VFL ? and whether there is conflict between quality predict branch and GFL/VFL?
I hope you can tell me the understanding of these questions. I would appreciate your reply！