Few-Shot Object Detection

(ECCV 2020) PyTorch implementation of paper "Few-Shot Object Detection and Viewpoint Estimation for Objects in the Wild"
[PDF] [Project webpage] [Code (Viewpoint)]

If our project is helpful for your research, please consider citing:

@INPROCEEDINGS{Xiao2020FSDetView,
    author    = {Yang Xiao and Renaud Marlet},
    title     = {Few-Shot Object Detection and Viewpoint Estimation for Objects in the Wild},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year      = {2020}}

ChangeLog

• [Dec-15 2020] There could be a problem for the COCO base model weight downloaded in download_models.sh, please download the correct model weight from here and put it in save_models/COCO/.

Tabel of Contents

• Installation
• Data Preparation
• Getting Started
• Quantitative Results

Installation

Code built on top of MetaR-CNN.

Requirements

• CUDA 8.0
• Python=3.6
• PyTorch=0.4.0
• torchvision=0.2.1
• gcc >= 4.9

Build

• Create conda env:

conda create --name FSdetection --file spec-file.txt
conda activate FSdetection

• Compile the CUDA dependencies:

cd {repo_root}/lib
sh make.sh

Data Preparation

We evaluate our method on two commonly-used benchmarks. See data/README.md for more details.

PASCAL VOC

We use the train/val sets of PASCAL VOC 2007+2012 for training and the test set of PASCAL VOC 2007 for evaluation. We split the 20 object classes into 15 base classes and 5 novel classes, and we consider 3 splits proposed in FSRW.

Download PASCAL VOC 2007+2012, create softlink named VOCdevkit in the folder data/.

COCO

We use COCO 2014 and keep the 5k images from minival set for evaluation and use the rest for training. We use the 20 object classes that are the same with PASCAL VOC as novel classes and use the rest as base classes.

Download COCO 2014, create softlink named coco in the folder data/. Please follow the instructins here to download the instances_minival2014.json and instances_valminusminival2014.json.

Getting Started

1. Base-Class Training

Pre-trained ResNet: folloing Meta R-CNN, we used ResNet101 for PASCAL VOC and ResNet50 for MS-COCO. Download it and put it into the data/pretrained_model/.

• We provide pre-trained models of base-class training:

bash download_models.sh

You will get a dir like:

save_models/
    COCO/
    VOC_first/
    VOC_second/
    VOC_third/

• You can also train it yourself:

# the first split on VOC
bash run/train_voc_first.sh

# the second split on VOC
bash run/train_voc_second.sh

# the third split on VOC
bash run/train_voc_third.sh

# NonVOC / VOC split on COCO
bash run/train_coco.sh

2. Few-Shot Fine-tuning

Fine-tune the base-training models on a balanced training data including both base and novel classes (3K instances per base class and K instances per novel class):

bash run/finetune_voc_first.sh

bash run/finetune_voc_second.sh

bash run/finetune_voc_third.sh

bash run/finetune_coco.sh

3. Testing

Evaluation is conducted on the test set of PASCAL VOC 2007 or minival set of COCO 2014:

bash run/test_voc_first.sh

bash run/test_voc_second.sh

bash run/test_voc_third.sh

bash run/test_coco.sh

Quantitative Results

Multiple Runs

By running multiple times (~10) the few-shot fine-tuning experiments and averaging the results, we got the performance below:

Pascal-VOC (AP@50)

MS-COCO

Specific Split

For a direct and quick comparison on COCO, we also run experiments using the specific few-shot sample split provided in TFA.

• Download their json files into the annotation folder of COCO:

cd ./data/coco/annotations
mkdir TFA && cd TFA
wget -r --no-parent  http://dl.yf.io/fs-det/datasets/cocosplit/
mv dl.yf.io/fs-det/datasets/cocosplit/ cocosplit && rm -r dl.yf.io 

• You will see a set of json files in format "full_box_{K}shot_{cls}_trainval.json" as well as 9 folders named "seed{i}"

• Then you can run the command line in run/finetune_coco_TFA.sh to finish few-shot fine-tuning and testing

MS-COCO

We get the following results:

Note: the difference between the performance of multiple runs and the performance of this specific split can be explained by the different sample configuration in the few-shot fine-tuning stage. As we follow the strategy proposed in MetaR-CNN to sample 3K instances of base class and K instances of novel class, without any specific adjustment, the performance would naturally degrade when we simply use the split of TFA where only K instances are considered for each base class.