VidStabFormer: Full-frame Video Stabilization via Spatial-Temporal Transformers

Official implementation of "VidStabFormer: Full-frame Video Stabilization via Spatial-Temporal Transformers"

Introduction

This study proposes a new full-frame video stabilization model leveraging spatial-temporal transformer architecture to fill missing regions that arise after the camera path smoothing. We also propose a new temporal self-supervised learning strategy to train the model. Experimental results show that the proposed model performs better than existing full-frame video stabilization counterparts.

Installation

• Test environment packages are in packages.txt. We will prepare it as yaml file soon!

• Download the pre-trained model and put it inside the "models" folder.

• Clone https://github.com/sniklaus/pytorch-pwc inside model folder. We only used a modified version of its backwarp function.

• Follow the CVPR2020 paper implementation of "Yu and Ramamoorthi, 2020" in https://github.com/alex04072000/FuSta/tree/main for creating warping field of any video.

• Put video and warping fields in the data folder. You can use "1" for video frames and "1_wf" for warping fields. You can put multiple videos. Add video frame numbers inside test.txt.

• Example video and related warping fields

Testing VidStabFormer

You can test VidStabFormer using the following script.

./test.sh

Contact

Please open an issue for any problem. We will answer it as soon as possible.

Citing VidStabFormer

@article{Karacan2025, 
author = {Levent Karacan and Mehmet Sarıgül},
title = {Full-frame video stabilization via spatiotemporal transformers},
year = {2025},
journal = {Computational Visual Media},
volume = {11},
number = {3},
pages = {655-667}
}