Transferable-guided Attention Is All You Need for Video Domain Adaptation

André Sacilotti¹   Samuel Felipe dos Santos²   Nicu Sebe³   Jurandy Almeida²  

¹University of São Paulo   ²Federal University of São Carlos   ³University of Trento

Accepted at WACV 2025

About

TransferAttn is a framework for unsupervised domain adaptation (UDA) in videos that leverages Vision Transformers (ViT) by incorporating spatial and temporal transferability into a attention mechanism (DTAB).

DTAB Overview

Dataset Preparation

We provide the extracted features used in all experiments. The datasets must be extracted in the folder "{root_path}/dataset"

Dataset	Google Drive
UCF101	Download I3D
UCF101	Download STAM
HMDB51	Download I3D
HMDB51	Download STAM
Kinetics and NEC-Drone	Download STAM*

The Kinetics-Gameplay dataset is a private dataset, so, we cannot share the extracted features. Please refer to TA3N to request access.

*We are having some trouble for uploading the features due to university Google Drive limitations. We aim to make avaible soon, sorry for the inconvinience.

Getting Started

First, you might need to create the conda environment as following:

conda env create --file enviroment.yml

After that, you can run the experiments:

bash run.sh

The experiments were run on the following GPUS: GTX 1080 Ti, and Titan X. Some variation in performance may occur on different hardwares due to architecture changes.

TODO List

• Code release.
• Release download links of extracted features.

Acknowledgement

We acknowledge the use of the following public resources in the development of this work: UCF₁₀₁,HMDB₅₁, NEC-Drone, UDAVT and TranSVAE. Special acknowledge for MA2LT-D who was the base of our code development.

Citation

If you find TransferAttn useful for your work please consider citing:

@InProceedings{WACV_2025_Sacilotti,
author = {A. {Sacilotti} and S. F. {Santos} and N. {Sebe} and J. {Almeida}},
title = {Transferable-guided Attention Is All You Need for Video Domain Adaptation},
pages = {1–11},
booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
address = {Tucson, AZ, USA},
month = {February 28 – March 4},
year = {2025},
publisher = {{IEEE}},
}

License

This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.