Awesome

Box-Level Active Detection (CVPR2023 highlight)

[Project Website] | [Paper] <br>

Introduction

This repo provides the official implementation of CVPR2023 paper Box-level Active Detection, with a unified codebase for active learning for detection.

Benchmark and model zoo

Supported datasets:

PASCAL VOC 0712
COCO Detection

Supported active learning methods:

Usage

Requirements

python=3.6
Pytorch=1.9.1
torchvision=0.10.1
mmcv=1.3.9
mmdetection=2.16.0

Installation

make install

And if you'd like to save the best checkpoint during training in mmdet 2.16, fix the Line 295 in mmcv/runner/hooks/evaluation.py as

runner.save_checkpoint(
                runner.work_dir, filename_tmpl=best_ckpt_name, create_symlink=False)

Dataset

Datasets are placed in data/detection/<dataset>, otherwise the data_root variable in config files should be updated.

Note that we convert the annotations of PASCAL VOC into COCO format with the dataset_converter.

Iterative Training and Evaluation

For example, run the following command for ComPAS on PASCAL VOC0712:

SEED=2022 QUERY_UNIT=box INIT_NUM=3000 ADD_NUM=1000 TRAIN_STEP=10 GPUS=4 bash dist_run_compas.sh voc0712 box_compas configs/mining/faster_rcnn/augs/faster_rcnn_r50_fpn_1x_voc0712_partial.py --deterministic

The results reported in the paper were conducted with seeds 2020, 2021, 2022.

Citation

If this toolbox or benchmark is useful in your research, please cite this project.

@InProceedings{blad2023,
    author    = {Lyu, Mengyao and Zhou, Jundong and Chen, Hui and Huang, Yijie and Yu, Dongdong and Li, Yaqian and Guo, Yandong and Guo, Yuchen and Xiang, Liuyu and Ding, Guiguang},
    title     = {Box-Level Active Detection},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {23766-23775}
}