Equalization Loss for Long-Tailed Object Recognition

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

:warning: 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]

<div align="center"> <img width="70%", src="https://tztztztztz.github.io/images/eql-gradient.jpg"/> </div><br/>

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

<table><tbody> <!-- START TABLE --> <!-- TABLE HEADER --> <th valign="bottom", align="left">Backbone</th> <th valign="bottom", align="left">Method</th> <th valign="bottom">AP</th> <th valign="bottom">AP.r</th> <th valign="bottom">AP.c</th> <th valign="bottom">AP.f</th> <th valign="bottom">AP.bbox</th> <th valign="bottom">download</th> <!-- TABLE BODY --> <tr> <td align="left">R50-FPN</td> <td align="left">MaskRCNN</td> <td align="center">21.2</td> <td align="center">3.2</td> <td align="center">21.1</td> <td align="center">28.7</td> <td align="center">20.8</td> <td align="center"><a href="https://www.dropbox.com/s/63q8cf7i62aveo6/model_final.pth?dl=0">model</a>&nbsp;|&nbsp;<a href="https://www.dropbox.com/s/wyfavlqzw17u8qf/log.txt?dl=0">metrics</a></td> </tr> <tr> <td align="left">R50-FPN</td> <td align="left">MaskRCNN-EQL</td> <td align="center">24.0</td> <td align="center">9.4</td> <td align="center">25.2</td> <td align="center">28.4</td> <td align="center">23.6</td> <td align="center"><a href="https://www.dropbox.com/s/ebto3pqdjb24u3x/model_final.pth?dl=0">model</a>&nbsp;|&nbsp;<a href="https://www.dropbox.com/s/fpyb0yfmaydnpva/log.txt?dl=0">metrics</a></td> </tr> <tr> <td align="left">R50-FPN</td> <td align="left">MaskRCNN-EQL-Resampling</td> <td align="center">26.1</td> <td align="center">17.2</td> <td align="center">27.3</td> <td align="center">28.2</td> <td align="center">25.4</td> <td align="center"><a href="https://www.dropbox.com/s/rwoq39ji4lywnh9/model_final.pth?dl=0">model</a>&nbsp;|&nbsp;<a href="https://www.dropbox.com/s/8m3pvp6lh0qyahd/log.txt?dl=0">metrics</a></td> </tr> <tr> <td align="left">R101-FPN</td> <td align="left">MaskRCNN</td> <td align="center">22.8</td> <td align="center">4.3</td> <td align="center">22.7</td> <td align="center">30.2</td> <td align="center">22.3</td> <td align="center"><a href="#">model</a>&nbsp;|&nbsp;<a href="https://www.dropbox.com/s/oe4wyqzbukl6p5r/log.txt?dl=0">metrics</a></td> </tr> <tr> <td align="left">R101-FPN</td> <td align="left">MaskRCNN-EQL</td> <td align="center">25.9</td> <td align="center">10.0</td> <td align="center">27.9</td> <td align="center">29.8</td> <td align="center">25.9</td> <td align="center"><a href="#">model</a>&nbsp;|&nbsp;<a href="https://www.dropbox.com/s/o5xr7csae3sf3or/log.txt?dl=0">metrics</a></td> </tr> <tr> <td align="left">R101-FPN</td> <td align="left">MaskRCNN-EQL-Resampling</td> <td align="center">27.4</td> <td align="center">17.3</td> <td align="center">29.0</td> <td align="center">29.4</td> <td align="center">27.1</td> <td align="center"><a href="#">model</a>&nbsp;|&nbsp;<a href="https://www.dropbox.com/s/303juk6ny9y6f0t/log.txt?dl=0">metrics</a></td> </tr> </tbody></table>

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.

<a name="CitingEQL"></a>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}
}