Asymmetric Event-Image Stereo with Temporal Feature Gating and Iterative Structure-Detail Refinement


Hao Zhuang (Beijing Institute of Technology), Yan Yang (Australian National University), Liyuan Pan (Beijing Institute of Technology)
The 35th British Machine Vision Conference

Abstract

This paper studies disparity estimation using asymmetric stereo inputs, combining event streams and conventional frame-based images. Past works directly calculate the similarity between features from the two modalities for making predictions, neglecting the inherent domain shift between them. First, event data are both temporally and spatially aware, while a frame is spatially aware. Second, event data are well-suited for capturing scene structure, whereas frame-based images excel at preserving fine-grained local visual details. To address these challenges, our method introduces two key components: (1) We enhance frame features by gating them with temporal cues extracted from consecutive frames, which improves their temporal awareness for more accurate similarity computation. (2) We perform disparity estimation using an iterative refinement scheme that alternates between updating structure and local features. This process starts with reliable structure features and progressively refining the disparity prediction by integrating fine detail. Extensive experiments on the MVSEC and DSEC datasets show that our method significantly outperforms prior state-of-the-art approaches.

Citation

@inproceedings{Zhuang_2025_BMVC,
author    = {Hao Zhuang and Yan Yang and Liyuan Pan},
title     = {Asymmetric Event-Image Stereo with Temporal Feature Gating and Iterative Structure-Detail Refinement},
booktitle = {36th British Machine Vision Conference 2025, {BMVC} 2025, Sheffield, UK, November 24-27, 2025},
publisher = {BMVA},
year      = {2025},
url       = {https://bmva-archive.org.uk/bmvc/2025/assets/papers/Paper_851/paper.pdf}
}


Copyright © 2025 The British Machine Vision Association and Society for Pattern Recognition
The British Machine Vision Conference is organised by The British Machine Vision Association and Society for Pattern Recognition. The Association is a Company limited by guarantee, No.2543446, and a non-profit-making body, registered in England and Wales as Charity No.1002307 (Registered Office: Dept. of Computer Science, Durham University, South Road, Durham, DH1 3LE, UK).

Imprint | Data Protection