Introduction

This repository contains the Pytorch implementation of our paper Regularizing Classifier For Few-Shot Object Detection via Feature Augmentation and Random Labeling. Our code is based on the baseline algorithm DeFRCN.

The methods of our paper are mainly implemented in FARL/defrcn/modeling/roi_heads/roi_heads.py.

Quick Start

1. Prepare Requirements

• Create a virtual environment

• Install PyTorch 1.6.0 with CUDA 10.1 (other versions may also work)

  pip3 install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

• Install Detectron2

  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

2. Prepare Data and Weights

• Data Preparation

  According to our baseline algorithm DeFRCN's data preparation, we can download the datasets using the following links:

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

  Unzip the downloaded datasets and put them into project directory as follows:

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

• Weights Preparation

  Our baseline algorithm DeFRCN designs a PCB module to improve the classification performance, which utilize a ImageNet pretrained model. The ImageNet pretrained model can be downloaded from GoogleDrive. Unzip the file, and the model named resnet101-5d3b4d8f.pth is for PCB module.

  Our paper targets the fine-tuning stage, so we need to download the weight after base training from our baseline algorithm, and use it as initial weight for the fine-tuning process. As for DeFRCN, we can download their results from GoogleDrive and GoogleDrive for PASCAL VOC and MSCOCO respectively. Unzip the files and rename the weights as follows:

  Path	Rename	Note
  voc/defrcn_one/defrcn_det_r101_base1/model_reset_surgery.pth	voc_split1_model_reset_surgery.pth	weight after base training for VOC split1
  voc/defrcn_one/defrcn_det_r101_base2/model_reset_surgery.pth	voc_split2_model_reset_surgery.pth	weight after base training for VOC split2
  voc/defrcn_one/defrcn_det_r101_base3/model_reset_surgery.pth	voc_split3_model_reset_surgery.pth	weight after base training for VOC split3
  coco/defrcn_one/defrcn_det_r101_base/model_reset_surgery.pth	coco_model_reset_surgery.pth	weight after base training for COCO

  Put the pretrained weights into project directory as follows:

  FARL
    | -- ...
    | -- pretrained
    |     | -- resnet101-5d3b4d8f.pth
    |     | -- voc_split1_model_reset_surgery.pth
    |     | -- voc_split2_model_reset_surgery.pth
    |     | -- voc_split3_model_reset_surgery.pth
    |     | -- coco_model_reset_surgery.pth
    | -- ...
  

3. Train and Evaluate

• PASCAL VOC

  bash run_voc.sh EXP_NAME SPLIT_ID (1, 2 or 3) METHOD (baseline, FA, FARL or iFARL)

• MSCOCO

  bash run_coco.sh EXP_NAME METHOD (baseline, FA, FARL or iFARL)