Need help with FSOD-code?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

204 Stars 53 Forks MIT License 31 Commits 54 Opened issues

Services available


Need anything else?

Contributors list

# 230,840
27 commits

A Pytorch Implementation of FSOD


Detectron2 based FSOD implementation is released in FewX! It can reach 12.0 AP on MS COCO (10-shot)!

Getting Started

Clone the repo:

git clone


Tested under python3.

  • python packages
    • pytorch==0.4.1
    • torchvision>=0.2.0
    • cython
    • matplotlib
    • numpy
    • scipy
    • opencv
    • pyyaml==3.12
    • packaging
    • pandas
    • pycocotools — for COCO dataset, also available from pip.
    • tensorboardX — for logging the losses in Tensorboard
  • An NVIDAI GPU and CUDA 9.0 are required. (Do not use other versions)
  • NOTICE: different versions of Pytorch package have different memory usages.


Compile the CUDA code:

cd lib  # please change to this directory

If your are using Volta GPUs, uncomment this line in

and remember to postpend a backslash at the line above.
defaults to
. If you want to use a CUDA library on different path, change this line accordingly.

It will compile all the modules you need, including NMS, ROIPooing, ROICrop and ROI_Align. (Actually gpu nms is never used ...)

Note that, If you use

to set gpus, make sure at least one gpu is visible when compile the code.

Data Preparation

Please add

in the
directory and the structure is :
      └── fsod
            ├── code files
            └── data
                  ├──── fsod
                  |       ├── annotations
                  │       │       ├── fsod_train.json
                  │       │       └── fsod_test.json
                  │       └── images
                  │             ├── part_1
                  │             └── part_2
                  └──── pretrained_model
                          └── model_final.pkl (from detectron model zoo: End-to-End Faster & Mask R-CNN Baselines R-50-C4 Faster 2x model)

You can download the model_final.pkl from here: Model link

Training and evaluation

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 tools/ --save_dir fsod_save_dir --dataset fsod --cfg configs/fsod/voc_e2e_faster_rcnn_R-50-C4_1x_old_1.yaml --bs 4 --iter_size 2 --nw 4 --load_detectron data/pretrained_model/model_final.pkl

CUDA_VISIBLE_DEVICES=0,1,2,3 python3 tools/ --multi-gpu-testing --dataset fsod --cfg configs/fsod/voc_e2e_faster_rcnn_R-50-C4_1x_old_1.yaml --load_ckpt Outputs/fsod_save_dir/ckpt/model_step59999.pth

The default setting is on 4 GPUs. If you want to change GPU numbers, please change

in the training script and make sure
. If you want to use the default training setting on 4 GPUs, you can directly run
to train and evaluate the model.


Please note that this repository is only for episode training and evaluation on FSOD dataset. Other expriments on different datasets are in different evaluation settings and I will gradually merge them. Currently, the code only supports 1 image on per GPU for both training and evaluation.

This repository is originally built on roytseng-tw/Detectron.pytorch. You can reference it for more implementation details.


If you use this dataset in your research, please cite this paper.

    title={Few-Shot Object Detection with Attention-RPN and Multi-Relation Detector},
    author={Fan, Qi and Zhuo, Wei and Tang, Chi-Keung and Tai, Yu-Wing},

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.