Home

Awesome

Sof-DETR

This repository contains PyTorch implemetation of the paper "Improving Small Objects Detection using Transformer"
If you find our paper or provided codes helpful in your research, then please do not forget to cite our paper. Thank you!
The following architecture represents our proposed model Sof-DETR for Object Detection.

alt text

Requirements

<pre> -python 3.8.8 -pytorch 1.8.1 -torchvision 0.9.1 -numpy 1.18.5 -cudatoolkit 11.1.74 -scipy 1.4.1 -tensorboard 2.4.0 -tensorflow-gpu 2.3.0 -requests 2.24.0 </pre>

Testing

We have provided the Jupyter Notebooks for better visualization of the predicted class and bounding boxes and classes.\ Our jupyter notebooks also display the self-attention maps and decoder feature maps.
Two Jupyter Notebooks are provided :

  1. test_sof-detr.ipynb
  2. sof-detr_attention.ipynb

Sof-DETR Evaluation on MSCOCO Detection Dataset 2017 (val set)

ModelAP-allAP-50AP-75AP-SmallAP-MediumAP-Large
SOF-DETR (Resnet-50)42.761.845.421.745.961.5

Sof-DETR Online Evaluation on MSCOCO Detection Dataset 2017 (test-dev set)

ModelAP-allAP-50AP-75AP-SmallAP-MediumAP-LargeAR-max=1AR-max=10AR-max=100AR-SmallAR-MediumAR-Large
SOF-DETR (Resnet-50)43.062.046.021.046.059.034.055.059.032.064.081.0

Citation

Please cite the following BibTex:

If you find the paper and this repository helpful, please consider citing our paper Sof-DETR. Thank you!

License

This project is licensed under Machine Learning & Vision Laboratory (MLV Lab), GIST.

Acknowledgments

We would like to thanks facebookresearch team.