csleizhang / Transfer-Learning-Library-for-Object-Detection

Transfer Learning Library for Domain Adaptation and Domain Generalization of Object Detection.

Transfer-Learning-Library-for-Object-Detection

Transfer Learning Library for Domain Adaptation and Domain Generalization of Object Detection.

Introduction

This is a Transfer Learning library for Object Detection task. It is mainly implemented by PyTorch.

We provide the implementation of many methods in the directory methods. They belong to several different learning setups, including:

• UDAOD (Unsupervised Domain Adaptive Object Detection)
• DGOD (Domain Generalized Object Detection)
• UniDAOD (Universal Domain Adaptive Object Detection)

Provided Methods

The currently provided methods include:

Unsupervised Domain Adaptive Object Detection

• DAF -Domain Adaptive Faster R-CNN for Object Detection in the Wild [CVPR 2018][local code][corresponding original code]
• MAF -Multi-adversarial Faster-RCNN for Unrestricted Object Detection [ICCV 2019][local code][corresponding original code]
• ATF -Domain Adaptive Object Detection via Asymmetric Tri-way Faster-RCNN [ECCV 2020][local code][corresponding original code]
• IDF -Exploring Implicit Domain-invariant Features for Domain Adaptive Object Detection [TCSVT 2023][local code][corresponding original code]
• PA-ATF -Partial Alignment for Object Detection in the Wild [TCSVT 2022][local code]
• PT-MAF -Multi-adversarial Faster-RCNN with Paradigm Teacher for Unrestricted Object Detection [IJCV 2022][local code]

Domain Generalized Object Detection

• MAD -Multi-view Adversarial Discriminator: Mine the Non-causal Factors for Object Detection in Unseen Domains [CVPR 2023][local code][corresponding original code]

Universal Domain Adaptive Object Detection

• US-DAF -Universal Domain Adaptive Object Detector [ACM MM 2022][local code][corresponding original code]

Usage

Data Preparation

All datasets are aranged in the format of PASCAL VOC as follows:

# cityscapes   
- cityscape
    - VOC2007
        - ImageSets  
        - JPEGImages  
        - Annotations  

Pretrained Model

We use two pretrained models, VGG16 and ResNet101. You can download these two models from:

• VGG16: Dropbox, VT Server, Baidu Netdisk
• ResNet101: Dropbox, VT Server, Baidu Netdisk

Environment Configuration

The pytorch environment includes python 3.6, pytorch 0.4.0, CUDA 8.0 or higher, torchvision 0.2.1, matplotlib 3.3.4, etc. See the requirements.txt file for details.
The following is a simple example of configuring an environment using conda.

# create environment
conda create --name torch0_4 python=3.6
# activate environment
conda activate torch0_4
# install pytorch
conda install pytorch=0.4.0 cuda80 -c pytorch
# install other requirements
pip install -r requirements.txt

The default version we provide is compiled with Python 3.6, so if you're using another version of Python, use the following command to compile it yourself:

# compile the cuda dependencies
cd lib
sh make.sh

Train and Test

You can find methods in the directory methods. In general, there are four main files in each method directory, which are:
_train.py - Training script for the model
_test.py - Test script for the model
_train.sh - Detailed model train run commands
_test.sh - Detailed model test run commands

Results

Unsupervised Domain Adaptive Object Detection

• Cityscapes (source) -> Foggy cityscapes (target) backbone: VGG16

	person	rider	car	truck	bus	train	motorcycle	bicycle	mAP
DAF	29.9	41.2	43.3	20.2	36.3	27.6	26.4	33.4	32.3
MAF	33.2	44.9	44.3	28.5	40.1	23.9	30.9	37.4	35.4
ATF	33.8	46.7	44.9	26.5	45.5	32.5	34.9	38.2	37.9
IDF	36.4	48.6	52.4	33.9	52.3	35.2	36.9	39.6	41.9
PA-ATF	34.3	45.6	52.1	28.7	47.5	49.4	33.7	37.4	41.1
PT-MAF	34.2	50.4	50.0	27.3	47.2	46.1	32.5	38.5	40.8

Domain Generalized Object Detection

• Training Dataset:Cityscapes, Test Dataset:Foggy cityscapes backbone: VGG16

	person	rider	car	truck	bus	train	motorcycle	bicycle	mAP
MAD	33.9	46.9	45.1	28.0	44.1	34.4	33.7	39.3	38.2

Universal Domain Adaptive Object Detection

• VOC -> Clipart (The two datasets share 10 categories) backbone: ResNet101

	bus	car	cat	chair	cow	diningtable	dog	horse	motorbike	person	mAP
US-DAF	31.1	43.2	13.9	39.4	58.7	27.4	12.6	44.3	55.4	61.3	38.7

Contact

If you have any problem about our code, feel free to contact us.

• Zhilong Zhang ([e-mail])
• Lei Zhang ([e-mail])

or describe it in Issues.

Citation

If you use this toolbox or benchmark in your research, please cite this project.

@misc{lib,
    author = {Zhilong Zhang, Lei Zhang},
    title = {Transfer-Learning-Library-for-Object-Detection},
    year = {2023},
    publisher = {GitHub},
    journal = {GitHub repository},
    howpublished = {\url{https://github.com/live-group/Transfer-Learning-Library-for-Object-Detection}},
}

Acknowledgment

The base object detector in this library is built upon faster-rcnn.pytorch, and we'd like to appreciate for their excellent works.