Introduction

This repo contains the source for "Prototype Distance Ratio Sampling for Generalised Few Shot Object Detection"

Installation

Datasets

To automatically set up mamba and download the datasets:

source install_scripts/install_base.sh
mamba create -y -n tfa python=3.8 && source ~/.bashrc && mamba activate tfa

Training and validation splits for COCO should be sent to the same directory named trainval2014, with val2014 being just a symlink to trainval2014.

Reminder that VOC is 4.3 GB, COCO is 20 GB, so plan accordingly.

Dependencies

Using install_scripts/install_detectrion2.sh should do the job. If it doesn't work, here are the manual steps below:

Downgrade numpy to avoid deprecated alias errors (np.bool etc), downgrade PIL

 pip install numpy==1.23.1
pip install pillow==9.5.0

Avoid recent bug with new setuptools compat aws-neuron/aws-neuron-sdk#893

pip3 install setuptools==69.5.1

Install pytorch, detecton2 and dependencies

pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 -i https://download.pytorch.org/whl/cu117
pip install git+https://github.com/facebookresearch/fvcore
git clone https://github.com/facebookresearch/detectron2.git
pip install opencv-python scikit-learn

(Optional) Download determinism patch by M.Yousef explained in https://github.com/facebookresearch/detectron2/issues/4723, will ensure reproducibility on the same machine

cd detectron2
# Download determinism patch
curl -O https://storage.googleapis.com/gfsod_exp_data_mirror/roi_align_determinism.patch
# Apply patch to installed detectron. Easier to do on source install since it does not rely on python environment path.
patch -p1 <roi_align_determinism.patch

Install detectron2. Note: there are no binaries available since FB stopped releasing them years ago, but the build system works pretty well since they have integrated detectron2 into another project.

python -m pip install -e .
cd ..

Sampling

To generate different base splits, run voc_sampler.sh and coco_sampler.sh. Novel annotations will be copied automatically from the ones used in previous works (splits 0-9) keep it fair.

By default, all generated splits will have different numbers from the old ones in use by other G-FSOD works (10-19 for COCO, 30-39 for VOC). To change which splits you're running, edit DATA_SPLIT_LIST in deployment_cfg.sh

Training and Evaluation

• This follows the same structure as DeFRCN with a few differences: the scripts will use pre-trained weights by default for the base classes, unless told otherwise in deployment.cfg

• To reproduce the results on VOC, EXP_NAME can be any string (e.g defrcn, or something) and SPLIT_ID must be 1 or 2 or 3 (we consider 3 random splits like other papers).

  bash run_voc.sh EXP_NAME SPLIT_ID (1, 2 or 3)
  

• To reproduce the results on COCO, EXP_NAME can be any string (e.g defrcn, or something)

  bash run_coco.sh EXP_NAME
  

• Don't forget to change IMAGENET_PRETRAIN* to the path to your downloaded pretrained ImageNet weights.

If you have less than 24GB VRAM at your disposal, edit run_voc.sh and run_coco.sh to append the SOLVER.IMS_PER_BATCH option.

Acknowledgements

This repo is developed based on DeFRCN, TFA and Detectron2. The training was originally run on ARCHIE-WeST.