• Linux with Python >= 3.6
• PyTorch >= 1.6 & torchvision that matches the PyTorch version.
• CUDA 10.1, 10.2, 11.0 (maybe even higher)
• GCC >= 4.9

• Clone Code
• Create a virtual environment (optional)

  conda create -n bsdet python=3.7
  source activate bsdet
  

• Install PyTorch with CUDA

  conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch
  

• Install Detectron2

  python3 -m pip install detectron2==0.3 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu110/torch1.7/index.html
  

• Install other requirements

  python3 -m pip install -r requirements.txt
  

• Refer to DeFRCN and TFA
• The final dataset file structure is as follows:

  ...
  datasets
    | -- coco
          | -- trainval2014/*.jpg
          | -- val2014/*.jpg
          | -- annotations/*.json
    | -- cocosplit
    | -- VOC2007
    | -- VOC2012
    | -- vocsplit
  defrcn
  tools
  ...
  

• BSDet config files:

  configs 
    | -- coco_CosSimNegHead
    | -- voc_CosSimNegHead
  

• VOC

  bash run_voc.sh
  

  • EXP_NAME: save directory
  • CONFIG_NAME: used config directory
  • SPLIT_ID: id of voc split
  • GPUS: number of used GPUs

• COCO

  bash run_coco.sh
  

  • EXP_NAME: save directory
  • CONFIG_NAME: used config directory
  • GPUS: number of used GPUs

• VOC

  bash test_voc.sh
  

  • CONFIG_NAME: used config directory
  • CHECKPOINT: used checkpoint
  • shot: number of novel class samples
  • SPLIT_ID: id of voc split
  • GPUS: number of used GPUs

• COCO

  bash test_coco.sh
  

  • CONFIG_NAME: used config directory
  • CHECKPOINT: used checkpoint
  • shot: number of novel class samples
  • GPUS: number of used GPUs

There are two main types of FSOD test, single test and multiple test. We adopt the single test (such as MPSR). It is worth noting that DeFRCN has two fine-tuning settings: one predicts only novel classes during fine-tuning (fsod), and the other predicts both base classes and novel classes (gfsod). To remain consistent with other existing works (such as MPSR), we use the gfsod setting. In addition, we only run novel class samples with seed 0, as it is consistent with the samples used in existing single-test works. The mAP of DeFRCN compared in the paper comes from DeFRCN (the result of seed 0 in the log file), therefore it is not consistent with the mAP reported in the DeFRCN paper. However, this is a fair comparison.

This repo is developed based on DeFRCN and Detectron2. Please check them for more details and features.