CASE: Cross-modal Semantic Anchoring Alignment and Structure Enhancement for Universal Domain Adaptation

Official PyTorch implementation of CASE, accepted by IEEE Transactions on Multimedia (TMM).

Overview

Universal Domain Adaptation (UniDA) transfers knowledge from a labeled source domain to an unlabeled target domain when label sets only partially overlap. CASE addresses this with two modules:

• CSA (Cross-modal Semantic Anchoring Alignment): builds a shared semantic anchor space and aligns cross-domain samples via JS-divergence-weighted contrastive learning.
• SSE (Semantic Structure Enhancement): aggregates anchor responses with K-Means clustering and feeds the enhanced representation to an All-in-One (AIO) classifier for unknown-class detection.

Backbone: frozen CLIP ViT-B/16 + lightweight MLP-Attention-MLP adapter (~3.15M trainable parameters).

Environment

conda create -n case python=3.10 -y
conda activate case
pip install -r requirements.txt
python -c "import nltk; nltk.download('wordnet')"

Download CLIP ViT-B/16 weights to ckpt/clip/ViT-B-16.pt (see OpenAI CLIP).

Data Preparation

1. Download benchmark datasets and place images following the list files under DataSets/.
2. Extract CLIP features:

python get_features.py --dataset OfficeHome
python get_features.py --dataset Office
python get_features.py --dataset VisDA
python get_features.py --dataset DomainNet

3. Build semantic anchors (once per domain, or use the provided labels_anchor/ files):

python build_anchors.py --dataset OfficeHome --domain Art --topk 2

Training

Single task:

python train.py --dataset OfficeHome --source Art --target Clipart

Run all domain pairs:

python train.py --dataset OfficeHome

Main hyperparameters :

Parameter	Default
epochs	80
batch size	36
learning rate	1e-3
anchor top-k	2
evaluation	average H-score over last 10 epochs

UniDA class splits: Office (10/10/11), OfficeHome (10/5/50), VisDA (6/6/6), DomainNet (150/50/145).

Project Structure

CASE-UniDA/
├── train.py           # training & evaluation
├── model.py           # CASE (CSA + SSE + AIO)
├── data.py            # dataset loaders & class splits
├── utils.py           # labels, templates, helpers
├── get_features.py    # CLIP feature extraction
├── build_anchors.py   # semantic anchor construction
├── labels_anchor/     # pre-built anchor vocabularies
└── DataSets/          # dataset index files

Citation

If you find this work useful, please cite:

@article{li2026case,
  title={Cross-modal Semantic Anchoring Alignment and Structure Enhancement for Universal Domain Adaptation},
  author={Li, Feijiang and Wang, Yuefeng and Qian, Yuhua and Wang, Jieting},
  journal={IEEE Transactions on Multimedia},
  year={2026}
}

License

This project is released for academic research use.