Long-tail Detection with Effective Class-Margins

Introduction

This is an official implementation of Long-tail Detection with Effective Class-Margins.

Long-tail Detection with Effective Class-Margins
Jang Hyun Cho and Philipp Krähenbühl
ECCV 2022 (oral)

Contact: janghyuncho7 [at] utexas.edu.

Installation

Requirements

• Python 3.6+
• PyTorch 1.8+
• torchvision 0.9+
• mmdet 2.14+
• mmcv 1.3+

We tested our codebase on mmdet 2.24.1, mmcv 1.5.1, PyTorch 1.11.0, torchvision 0.12.0, and python 3.9.

Setup

To setup the code, please follow the commands below:

# Clone the repo.
git clone [email protected]:janghyuncho/ECM-Loss.git
cd ECM-Loss 

# Create conda env.
conda create --name ecm_loss python=3.8 -y 
conda activate ecm_loss

# Install PyTorch.
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

# Install mmcv.
pip install -U openmim
mim install mmcv-full

# And mmdetection. 
pip install mmdet 

# Install mmdet dependencies.
pip install -e .

# Additionally, install lvis-api. 
pip install lvis

Dataset

Please download LVIS dataset, and structure the folders as following.

data
  ├── lvis_v1
  |   ├── annotations
  │   │   │   ├── lvis_v1_val.json
  │   │   │   ├── lvis_v1_train.json
  │   ├── train2017
  │   │   ├── 000000004134.png
  │   │   ├── 000000031817.png
  │   │   ├── ......
  │   ├── val2017
  │   ├── test2017

Training with ECM-Loss

All training commands for our models can be found here. For example, you can train mask-rcnn with resnet-50 backbone for 12 epochs with the following command:

./sh_files/ecm_loss/r50_1x.sh 

ECM Loss is implemented here.

Pretrained Models on LVIS v1

The Google Drive links to pretrained weights and their config files.

Framework	Backbone	Schedule	Box AP	Mask AP	Weight	Config
Mask R-CNN	R50	1x	26.9	26.4	weight	config
Mask R-CNN	R50	2x	27.9	27.5	weight	config
Mask R-CNN	R101	2x	29.4	28.7	weight	config
Cascade Mask R-CNN	R101	2x	33.4	30.6	weight	config

Citation

If you use use ECM Loss, please cite our paper:

@inproceedings{hyun2022long,
  title={Long-tail detection with effective class-margins},
  author={Hyun Cho, Jang and Kr{\"a}henb{\"u}hl, Philipp},
  booktitle={European Conference on Computer Vision},
  pages={698--714},
  year={2022},
  organization={Springer}
}

Acknowledgement

ECM Loss is based on MMDetection.