🧦 SoC: Semantic Orthogonality Calibration for Test-Time Prompt Tuning

This is the official implementation for SoC: Semantic Orthogonal Calibration for Test-Time Prompt Tuning, CVPR 2026.

Getting started

1. Clone this repository.
2. Download the desired datasets, and the associated data_splits. The TPT repository gives links to download everything. Put the downloaded .json split files in data/data_splits
3. Setup the environment.

conda env create -f environment.yml
conda activate soc

Running experiments

Running baseline experiments.

# TPT on DTD using ViT-L/14
bash scripts/test_fg.sh DTD ViT-L/14 tpt
# O-TPT on DTD using ViT-L/14
bash scripts/test_fg.sh DTD ViT-L/14 tpt_otpt 18

Running our proposed SoC.

bash scripts/test_fg.sh DTD ViT-L/14 tpt_otpt 30

Citation

If you find this repository useful, please consider citing this paper:

@inproceedings{fillioux2026soc,
    author    = {Fillioux, Leo and Chakraborty, Omprakash and Ben Ayed, Ismail and Cournède, Paul-Henry and Christodoulidis, Stergios and  Vakalopoulou, Maria and Dolz, Jose},
    title     = {{SoC}: Semantic Orthogonal Calibration for Test-Time Prompt Tuning},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2026},
}