/SPE

Official implementation of ECCV2022 paper End-to-End Weakly Supervised Object Detection with Sparse Proposal Evolution

Primary LanguagePythonApache License 2.0Apache-2.0

End-to-End Weakly Supervised Object Detection with Sparse Proposal Evolution

Install

pip install torch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0
pip install timm
pip install opencv-python
pip install scipy
pip install pycocotools

Data Preperation

We support COCO format. For VOC format annotations, please transfer them to coco format first.

For example, in experiments on VOC0712 trainval first transfer annotations of VOC2007 trainval and VOC2012 trainval to coco format in a single json file. # note: to evaluate mAP performance, please use the original annotation of voc07 test.

Models

VOC0712: mAP=51.0

COCO2017: AP=7.9, AP50=19.5, AP75=5.5

Train & Val

COCO 2017:

python scripts/run_coco17.py

VOC 07+12

python scripts/run_voc0712.py