This repository contains the official implementation for the paper Efficient Few-Shot Object Detection via Knowledge Inheritance (TIP 2022). The codes are built upon Detectron2. Many thanks to their awesome open-source library.
We build a highly efficient pretrain-transfer framework (PTF) baseline without extra overheads for FSOD. Upon this baseline, we devise an initializer named knowledge inheritance (KI) to reliably initialize the novel weights for the box classifier. Qualitative results on Pascal-VOC, COCO and LVIS demonstrate our approach sets new state of the art with higher efficiency. We hope to motivate a trend toward powerful yet efficient few-shot technique development.
# install pytorch and cuda, please follow the versions specified in the command below
conda install pytorch==1.4.0 python=3.7 torchvision==0.5.0 cudatoolkit=10.0 -c pytorch
# clone our repo and install detectron2
git clone https://github.com/Ze-Yang/Efficient-FSOD.git
cd Efficient-FSOD && pip install -e .
# install pycocotools, lvis-api and tensorboard
pip install cython; pip install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'
pip install git+https://github.com/lvis-dataset/lvis-api.git
pip install tensorboard
To rebuild detectron2, rm -rf build/ **/*.so then pip install -e ..
You often need to rebuild detectron2 after reinstalling PyTorch.
To train a model, run
python tools/train_net.py --num-gpus 4 \
--config-file configs/COCO-detection/faster_rcnn_R_101_FPN_base.yaml
To evaluate the trained models, run
python tools/test_net.py --num-gpus 4 \
--config-file configs/COCO-detection/faster_rcnn_R_101_FPN_ours_10shot.yaml \
--eval-only
If you find this repository useful in your research, please consider to cite our paper with the following BibTeX entry.
@article{yang2022efficient,
title={Efficient few-shot object detection via knowledge inheritance},
author={Yang, Ze and Zhang, Chi and Li, Ruibo and Xu, Yi and Lin, Guosheng},
journal={IEEE Transactions on Image Processing},
volume={32},
pages={321--334},
year={2022},
publisher={IEEE}
}This repository is released under the Apache 2.0 license.