Introduction

This repo contains the official PyTorch implementation of D\&R

Quick Start

1. Check Requirements

• Linux with Python >= 3.6
• PyTorch >= 1.6 & torchvision that matches the PyTorch version.
• CUDA 10.1, 10.2
• GCC >= 4.9

2. Build

• Create a virtual environment (optional)

  conda create -n dandr python=3.7
  conda activate dandrzq

• Install PyTorch according to your CUDA version

• Install Detectron2 (the version of Detectron2 must be 0.3)

  python3 -m pip install detectron2==0.3 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu101/torch1.6/index.html

• Install other requirements.

  python3 -m pip install -r requirements.txt

3. Prepare Data and Weights

• Data Preparation (from DeFRCN)

  Dataset	Size	GoogleDrive	BaiduYun	Note
  VOC2007	0.8G	download	download	-
  VOC2012	3.5G	download	download	-
  vocsplit	<1M	download	download	refer from TFA
  COCO	~19G	-	-	download from offical
  cocosplit	174M	download	download	refer from TFA

  • Unzip the downloaded data-source to datasets and put it into your project directory:

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

• Weights Preparation

  • DeFRCN use the imagenet pretrain weights to initialize the model. Download the same models from (given by DeFRCN): GoogleDrive BaiduYun
  • Put the chekpoints into ImageNetPretrained/MSRA/R-101.pkl, ImageNetPretrained/torchvision, respectively

  • We provide the BASEWEIGHT (refer to run*.sh) we used.	Dataset	Split	Size	GoogleDrive
    VOC2007	1	203.8M	download
    VOC2007	2	203.8M	download
    VOC2007	3	203.8M	download
    COCO	-	206.2MB	download

• Text Embeddings Preparation

  • Refer to the official implementation of CLIP for text embedding generation.
  • Put the generated text embeddings into 'dataset/clip'

4. Training and Evaluation

• To reproduce the results on VOC,

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

• To reproduce the results on COCO

  sh run_coco.sh

• Please read the details of few-shot object detection pipeline in run_*.sh.

Acknowledgement

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