Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (DPA)

Code implementation for Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment Zhi Hu (知乎)

Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment<br> Yuanfan Zheng^1,2, Jinlin Wu^1,2, Wuyang Li³, Zhen Chen*¹<br>¹ CAIR, HKISI-CAS; ² MAIS, Institute of Automation, Chinese Academy of Sciences; ³ The Chinese University of Hong Kong

<div align="center"> <img src="src\framework.png" alt="Framework" style="width: 100%; max-width: 450px;"/> </div>

Requirements

• Ubuntu 18.04.5 LTS
• Python 3.6
• CUDA 10.0
• PyTorch 1.0.0
• Faster R-CNN

Compile the code

#Compile the cuda dependencies using following simple commands following [Faster R-CNN](https://github.com/jwyang/faster-rcnn.pytorch/tree/pytorch-1.0):
cd lib
python setup.py build develop

Pre-trained Models

• ResNet101: Dropbox VT Server

Components

Global-level Domain Private Alignment (GDPA)

• Path: UniDAOD-DPA/lib/model/utils
• Function: global_alignment()

Instance-level Domain Shared Alignment (IDSA)

• Path: UniDAOD-DPA/lib/model/da_faster_rcnn/
• File: openset_weight.py

Private Class Constraint (PCC)

• Path: UniDAOD-DPA/lib/model/utils
• Function: instance_alignment_private

Training and Test

Train the model

CUDA_VISIBLE_DEVICES=0 python -u da_train_net.py \
--max_epochs 10 --cuda --dataset voc2clipart_0.25 \
--net res101 --save_dir ./weight_model/voc2clipart_0.25 \
--pretrained_path XXXX/pretrained_model/resnet101_caffe.pth \
--gc --lc --da_use_contex --weight_consis 0.1 --lr_bound 0.1 --gmm_split 0.03

Test the well-trained model:

python test_clipart_0.25.py >> test-voc025.out

Train the model and test the well-trained model through the script:

sh train_scripts\train_voc2clipart_0.25.sh

If you have any questions , please contact me at 478756030@qq.com

@article{zheng2024universal,
  title={Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment},
  author={Zheng, Yuanfan and Wu, Jinlin and Li, Wuyang and Chen, Zhen},
  journal={arXiv preprint arXiv:2412.11443},
  year={2024}
}