LEA

<h1 align="center">Learning Endogenous Attention for Incremental Object Detection</h1>

Official codebase for the CVPR 2025 paper: Learning Endogenous Attention for Incremental Object Detection (📄 Paper). This repository is implemented based on Detectron2.

Contents

• Introduction
• Installation
• Dataset Preparation
• Running Experiments
• Citation

Introduction

Learning Endogenous Attention (LEA) provides an effective framework for incremental object detection, supporting continual learning on large-scale benchmarks such as COCO and VOC. If you find this repository helpful, please consider citing our work!

Installation

Follow these steps to set up the LEA environment:

# Create conda environment
conda create -n lea python=3.9
conda activate lea

# Install PyTorch and dependencies (CUDA 11.6)
conda install pytorch==1.12.1 torchvision==0.13.1 \
    torchaudio==0.12.1 cudatoolkit=11.6 \
    -c pytorch -c conda-forge

# Clone this repository
git clone https://github.com/SONGX1997/LEA.git
cd LEA

# Install LEA (and Detectron2 dependencies)
python -m pip install -e LEA

For more details, refer to the Detectron2 Installation Guide.

Dataset Preparation

Prepare datasets for training and evaluation:

# Download COCO 2017 and split the data
python datasets/coco_deal.py

Running Experiments

Train and evaluate LEA on COCO datasets. You can adjust settings in the provided script:

bash run_coco.sh

Citation

If you use LEA, please cite:

@inproceedings{LEA_CVPR2025,
  title     = {Learning Endogenous Attention for Incremental Object Detection},
  author    = {X. Song and Y. He and J. Li and Q. Wang and Y. Gong},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2025},
}