Context-Transformer: Tackling Object Confusion for Few-Shot Detection

This repository contains the official implementation of the AAAI 2020 paper Context-Transformer: Tackling Object Confusion for Few-Shot Detection.

Introduction

To tackle the object confusion problem in few-shot detection, we propose a novel Context-Transformer within a concise deep transfer framework. Specifically, Context-Transformer can effectively leverage source-domain object knowledge as guidance, and automatically formulate relational context clues to enhance the detector's generalization capcity to the target domain. It can be flexibly embedded in the popular SSD-style detectors, which makes it a plug-and-play module for end-to-end few-shot learning. For more details, please refer to our original paper.

Transfer Setting COCO60 to VOC20 (Novel Class mAP)

Method	1shot	5shot
Prototype	22.8	39.8
Imprinted	24.5	40.9
Non-local	25.2	41.0
Baseline	21.5	39.4
Ours	27.0	43.8

News: We now support instance shot for COCO60 to VOC20 transfer setting, denoted by suffix -IS below.

Method	1shot	5shot
Baseline-IS	19.2	35.7
Ours-IS	27.1	40.4

Note:

• The instance shots are kept the same as incremental setting, which is different from the image shots we originally used in transfer setting. Therefore, it's possible that the 1-shot result of Ours-IS (27.1) is comparable to Ours (27.0).

Incremental Setting VOC15 to VOC20 (Novel Class mAP)

Method (1-shot)	Split1	Split2	Split3
Shmelkov2017	23.9	19.2	21.4
Kang2019	14.8	15.7	19.2
Ours	39.8	32.5	34.0

Method (5-shot)	Split1	Split2	Split3
Shmelkov2017	38.8	32.5	31.8
Kang2019	33.9	30.1	40.6
Ours	44.2	36.3	40.8

Note:

• The results here is higher than that reported in the paper due to training strategy adjustment.

License

Context-Transformer is released under the MIT License (refer to the LICENSE file for details).

Citing Context-Transformer

If you find Context-Transformer useful in your research, please consider citing:

@inproceedings{yang2020context,
  title={Context-Transformer: Tackling Object Confusion for Few-Shot Detection.},
  author={Yang, Ze and Wang, Yali and Chen, Xianyu and Liu, Jianzhuang and Qiao, Yu},
  booktitle={AAAI},
  pages={12653--12660},
  year={2020}
}

 

Contents

1. Installation
2. Datasets
3. Training
4. Evaluation

Installation

• Clone this repository. This repository is mainly based on RFBNet and Detectron2, many thanks to them.

• Install anaconda and requirements:

  • python 3.6

  • PyTorch 1.4.0

  • CUDA 10.0

  • gcc 5.4

  • cython

  • opencv

  • matplotlib

  • tabulate

  • termcolor

  • tensorboard

    You can setup the entire environment simply using conda:

    conda create -n CT python=3.6 && conda activate CT
    conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
    conda install cython opencv matplotlib tabulate termcolor tensorboard

• Compile the nms and coco tools:

sh make.sh

Note:

• Check your GPU architecture support in utils/build.py, line 131. Default is:

'nvcc': ['-arch=sm_61',

• Ensure that the cuda environment is integrally installed, including compiler, tools and libraries. Plus, make sure the cudatoolkit version in the conda environment matches with the one you compile with. Check about that using nvcc -V and conda list | grep cudatoolkit, the output version should be the same.
• We have test the code on PyTorch-1.4.0 and Python 3.6. It might be able to run on other versions but with no guarantee.

Datasets

VOC Dataset

Download VOC2007 trainval & test

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2007.sh # <directory>

Download VOC2012 trainval

# specify a directory for dataset to be downloaded into, else default is ~/data/
sh data/scripts/VOC2012.sh # <directory>

Create symlink for the VOC dataset:

ln -s /path/to/VOCdevkit data/VOCdevkit

Image shots and splits preparation

Move the Main2007.zip and Main2012.zip under data/ folder to data/VOCdevkit/VOC2007/ImageSets/ and data/VOCdevkit/VOC2012/ImageSets/ respectively, and unzip them. Make sure that the .txt files contained in the zip file are under corresponding path/to/Main/ folder.

COCO Dataset

Download COCO benchmark

Download the MS COCO dataset from official website to data/COCO/ (or make a symlink ln -s /path/to/coco data/COCO). All annotation files (.json) should be placed under the COCO/annotations/ folder. It should have this basic structure

$COCO/
$COCO/cache/
$COCO/annotations/
$COCO/images/
$COCO/images/train2014/
$COCO/images/val2014/

Note: The current COCO dataset has released new train2017 and val2017 sets which are just new splits of the same image sets.

Image splits preparation

Run the following command to obtain nonvoc/voc split annotation files (.json):

python data/split_coco_dataset_voc_nonvoc.py

Training

First download the fc-reduced VGG-16 PyTorch base network weights at https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth or from BaiduYun Driver, and place it under the directory weights/.

Phase 1

Transfer Setting

To pretrain RFBNet on source domain dataset COCO60:

python train.py --save-folder weights/COCO60_pretrain -d COCO -p 1

Incremental Setting

To pretrain RFBNet on VOC split1 (simply change --split for other splits):

python train.py --save-folder weights/VOC_split1_pretrain -d VOC -p 1 -max 50000 --steps 30000 40000 --checkpoint-period 4000 --warmup-iter 1000 --setting incre --split 1

Note:

• To ease your reproduce, feel free to download the above pretrained RFBNet models via BaiduYun Driver or OneDrive directly.

Phase 2

Transfer Setting

To finetune on VOC dataset (1 shot):

python train.py --load-file weights/COCO60_pretrain/model_final.pth --save-folder weights/fewshot/transfer/VOC_1shot -d VOC -p 2 --shot 1 --method ours -max 2000 --steps 1500 1750 --checkpoint-period 200 --warmup-iter 0 --no-mixup-iter 750 -b 20

To finetune on VOC dataset (5 shot):

python train.py --load-file weights/COCO60_pretrain/model_final.pth --save-folder weights/fewshot/transfer/VOC_5shot -d VOC -p 2 --shot 5 --method ours -max 4000 --steps 3000 3500 --checkpoint-period 500 --warmup-iter 0 --no-mixup-iter 1500

Incremental Setting

To finetune on VOC dataset for split1 setting (1 shot):

python train.py -d VOC --split 1 --setting incre -p 2 -m ours --shot 1 --save-folder weights/fewshot/incre/VOC_split1_1shot --load-file weights/VOC_split1_pretrain/model_final.pth -max 200 --steps 150 --checkpoint-period 50 --warmup-iter 0 --no-mixup-iter 100

To finetune on VOC dataset for split1 setting (5 shot):

python train.py -d VOC --split 1 --setting incre -p 2 -m ours --shot 5 --save-folder weights/fewshot/incre/VOC_split1_5shot --load-file weights/VOC_split1_pretrain/model_final.pth -max 400 --steps 350 --checkpoint-period 50 --warmup-iter 0 --no-mixup-iter 100

Note:

• Simply change --split for other split settings.
• For other shot settings, feel free to adjust --shot, -max, --steps and --no-mixup-iter to obtain satisfactory results.

Evaluation

Phase 1

Transfer Setting

To evaluate the pretrained model on COCO minival set:

python test.py -d COCO -p 1 --save-folder weights/COCO60_pretrain --resume

Incremental setting

To evaluate the pretrained model on VOC2007 test set (specify your target split via --split):

python test.py -d VOC --split 1 --setting incre -p 1 --save-folder weights/VOC_split1_pretrain --resume

Phase 2

Transfer Setting

To evaluate the transferred model on VOC2007 test set:

python test.py -d VOC -p 2 --save-folder weights/fewshot/transfer/VOC_5shot --resume

Incremental setting

To evaluate the incremental model on VOC2007 test set (specify your target split via --split):

python test.py -d VOC --split 1 --setting incre -p 2 --save-folder weights/fewshot/incre/VOC_split1_5shot --resume

Note:

• --resume: load model from the last checkpoint in the folder --save-folder.

If you would like to manually specify the path to load model, use --load-file path/to/model.pth instead of --resume.

 

Should you have any questions regarding this repo, feel free to email me at [email protected].