Equalization Loss for Long-Tailed Object Recognition

Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, Junjie Yan

⚠️ We recommend to use the EQLv2 repository (code) which is based on mmdetection. It also includes EQL and other algorithms, such as cRT (classifier-retraining), BAGS (BalanceGroup Softmax).

[arXiv] [BibTeX]

In this repository, we release code for Equalization Loss (EQL) in Detectron2. EQL protects the learning for rare categories from being at a disadvantage during the network parameter updating under the long-tailed situation.

Installation

Install Detectron 2 following INSTALL.md. You are ready to go!

LVIS Dataset

Following the instruction of README.md to set up the lvis dataset.

Training

To train a model with 8 GPUs run:

cd /path/to/detectron2/projects/EQL
python train_net.py --config-file configs/eql_mask_rcnn_R_50_FPN_1x.yaml --num-gpus 8

Evaluation

Model evaluation can be done similarly:

cd /path/to/detectron2/projects/EQL
python train_net.py --config-file configs/eql_mask_rcnn_R_50_FPN_1x.yaml --eval-only MODEL.WEIGHTS /path/to/model_checkpoint

Pretrained Models

Instance Segmentation on LVIS

Backbone	Method	AP	AP.r	AP.c	AP.f	AP.bbox	download
R50-FPN	MaskRCNN	21.2	3.2	21.1	28.7	20.8	model \| metrics
R50-FPN	MaskRCNN-EQL	24.0	9.4	25.2	28.4	23.6	model \| metrics
R50-FPN	MaskRCNN-EQL-Resampling	26.1	17.2	27.3	28.2	25.4	model \| metrics
R101-FPN	MaskRCNN	22.8	4.3	22.7	30.2	22.3	model \| metrics
R101-FPN	MaskRCNN-EQL	25.9	10.0	27.9	29.8	25.9	model \| metrics
R101-FPN	MaskRCNN-EQL-Resampling	27.4	17.3	29.0	29.4	27.1	model \| metrics

The AP in this repository is higher than that of the origin paper. Because all those models use:

Scale jitter
Class-specific mask head
Better ImageNet pretrain models (of caffe rather than pytorch)

Note that the final results of these configs have large variance across different runs.

Citing EQL

If you use EQL, please use the following BibTeX entry.

@InProceedings{tan2020eql,
  title={Equalization Loss for Long-Tailed Object Recognition},
  author={Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, 
  Wanli Ouyang, Changqing Yin, Junjie Yan},
  journal={ArXiv:2003.05176},
  year={2020}
}

TalentedMUSE/eql.detectron2