CFFlow: An Optical Flow Estimation Hinging on Cross-Frequency Attention


Mengfei Wang (Shanghai Institute of Microsystem and Information Technology), Dongchen Zhu (Shanghai Institute of Microsystem and Information Technology), Lei Wang (Shanghai Institute of Microsystem and Information Technology), Jiamao Li (Shanghai Institute of Microsystem and Information Technology)
The 35th British Machine Vision Conference
PDFPosterVideo (Right click to download)

Abstract

Due to issues in training datasets and motion estimation in optical flow, existing optical flow models overly emphasize surface-level information features, leading to inaccurate optical flow estimations. Through theoretical analysis, we reveal that multi-frame approaches require more attention to interconnections among diverse information than two-frame methods. Consequently, we propose \textbf{CFFlow}, which integrates cross-frequency information into optical flow estimation networks. To address challenges posed by large displacements and small object motion in real-world scenarios, we introduce two specialized modules: Large Displacement Attention (LDA) and Small Object Displacement Attention (SODA), designed to handle distinct motion patterns. CFFlow effectively resolves small target mismatches and large displacement challenges, achieving superior optical flow estimation accuracy. On Sintel and KITTI benchmarks, our CFFlow attains an Average Endpoint Error (AEPE) of \textbf{0.99} (clean pass) and \textbf{1.65} (final pass), with an F1-all error of \textbf{4.21\%}, ranking first among all three-frame and two-frame methods.

Citation

@inproceedings{Wang_2025_BMVC,
author    = {Mengfei Wang and Dongchen Zhu and Lei Wang and Jiamao Li},
title     = {CFFlow: An Optical Flow Estimation Hinging on Cross-Frequency Attention},
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_277/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