Pytorch implementation of our paper "IDa-Det: An Information Discrepancy-aware Distillation for 1-bit Detectors" accepted by ECCV2022.
Any problem, please contact the first author (Email: [email protected]).
Our code is heavily borrowed from DeFeat (https://github.com/ggjy/DeFeat.pytorch/) and based on MMDetection (https://github.com/open-mmlab/mmdetection).
- Python 3.7
- MMDetection 2.x
- This repo uses:
mmdet-v2.0
mmcv-0.5.6
cuda 10.1
- sh script.sh
We simplify and optimize the code. Now IDa-Det is successfully plugged in the original DeFeat project. The training cost is reduced by about 30% compared with the old version.
Pretrained model is here: GoogleDrive
Notes:
- Faster RCNN based model
- Batch: sample_per_gpu x gpu_num
Model | Batch | Lr schd | box AP | Model | Log |
---|---|---|---|---|---|
R101 | 4x2 | 0.01 | 81.9 | GoogleDrive | |
R101-BiR18 | 4x1 | 0.004 | 76.9 | GoogleDrive |
If you find this work useful in your research, please consider to cite:
@inproceedings{xu2022ida,
title={IDa-Det: An Information Discrepancy-Aware Distillation for 1-Bit Detectors},
author={Xu, Sheng and Li, Yanjing and Zeng, Bohan and Ma, Teli and Zhang, Baochang and Cao, Xianbin and Gao, Peng and L{\"u}, Jinhu},
booktitle={European Conference on Computer Vision},
pages={346--361},
year={2022},
organization={Springer}
}