by Yudong Liu, Tingting Liang, Yongtao Wang.
We have released our code at https://github.com/PKUbahuangliuhe/CBNet implemented by caffe2. Since Detectron code will not be maintained, we release our implementation based on mmdetection
Please follow mmdetection on how to install the environment.
You need to convert the original backbone to cbnet version with python convert_db.py or python convert_tb.py.
Our CBNetv2 will be released soon.
Contact us with bahuangliuhe@pku.edu.cn, tingtingliang@pku.edu.cn, wyt@pku.edu.cn.
The project is only free for academic research purposes, but needs authorization for commerce. For commerce permission, please contact wyt@pku.edu.cn.
If you use our code/model/data, please cite our paper: https://aaai.org/Papers/AAAI/2020GB/AAAI-LiuY.1833.pdf
| Baseline | Backbone | Input size | box AP |
|---|---|---|---|
| FPN | ResNext-101-32x4d | 1333x800 | 40.1 |
| FPN | Dual-ResNeXt-101-32x4d | 1333x800 | 41.5 |
| FPN | Triple-ResNeXt-101-32x4d | 1333x800 | 42.0 |
| FPN | Dual-ResNeXt-101-32x4d(CBNetv2) | 1333x800 | 43.2 |
| Cascade R-CNN(with DCN and multi-scale training) | Dual-ResNeXt-101-32x4d(CBNetv2) | 1333x800 | 51.2(51.6 on 2017test-dev) |