This is the implementation of CVPR 2020 paper "Rethinking Classification and Localization for Object Detection". The code is based on the maskrcnn-benchmark.
If the paper and code helps you, we would appreciate your kindly citations of our paper.
@inproceedings{wu2020rethinking,
title={Rethinking Classification and Localization for Object Detection},
author={Wu, Yue and Chen, Yinpeng and Yuan, Lu and Liu, Zicheng and Wang, Lijuan and Li, Hongzhi and Fu, Yun},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
year={2020}
}
Follow the maskrcnn-benchmark to install code and set up the dataset.
A docker image is also provided
docker pull yuewudocker/pytorchdoubleheads
If you use this docker, you can run the ./cmd_install.sh script for the installation.
Most experiments are done under the following environments:
PyTorch version: 1.0.0
OS: Ubuntu 16.04.3 LTS
Python version: 3.6
CUDA runtime version: 9.0.176
Nvidia driver version: 410.78
GPU: 4x Tesla P100-PCIE-16GB
Results on the COCO 2017 validation set:
| Backbone | AP | AP_0.5 | AP_0.7 | AP_s | AP_m | AP_l | Link |
|---|---|---|---|---|---|---|---|
| ResNet-50-FPN | 40.3 | 60.3 | 44.2 | 22.4 | 43.3 | 54.3 | model |
| ResNet-101-FPN | 41.9 | 62.4 | 45.9 | 23.9 | 45.2 | 55.8 | model |
Results on COCO 2017 test-dev:
| Backbone | AP | AP_0.5 | AP_0.7 | AP_s | AP_m | AP_l | Link |
|---|---|---|---|---|---|---|---|
| ResNet-101-FPN | 42.3 | 62.8 | 46.3 | 23.9 | 44.9 | 54.3 | bbox |
Use config files in ./configs/double_heads/ for Training and Testing.
Download models to the ./models directory. Then use the following script:
sh cmd_test.sh
You need modify the data path:
export DATA_DIR=/path/to/datafolder/
You can use the ./cmd_train.sh script to train with 4 gpus.
You have to modify following paths:
export OUTPUT_DIR=/path/to/modelfolder/
export PRETRAIN_MODEL=/path/to/pretrained/model
export DATA_DIR=/path/to/datafolder/