MAFNet:Multi-frequency Adaptive Fusion Network for Real-time Stereo Matching

2025-12-03

AI Summary

Existing stereo matching networks suffer from poor real-time performance and weak non-local contextual modeling on mobile devices. This paper proposes a multi-frequency adaptive fusion network based exclusively on efficient 2D convolutions. Our method addresses these challenges through three key contributions: (1) a frequency-domain adaptive filtering attention module that explicitly decomposes and fuses high- and low-frequency features; (2) integration of the Linformer low-rank attention mechanism to enable adaptive cross-frequency feature aggregation; and (3) a fully 2D convolutional architecture for cost volume construction and regularization, balancing computational efficiency with multi-scale representational capacity. Evaluated on Scene Flow and KITTI 2015 benchmarks, our approach significantly outperforms existing real-time stereo methods, achieving state-of-the-art trade-offs between accuracy and inference speed. The lightweight, convolution-only design ensures strong potential for deployment on resource-constrained mobile platforms.

Technology Category

• Computer Vision: Multi-modal Vision
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Machine Learning: Mixture of Experts (MoE)

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines

Abstract

Existing stereo matching networks typically rely on either cost-volume construction based on 3D convolutions or deformation methods based on iterative optimization. The former incurs significant computational overhead during cost aggregation, whereas the latter often lacks the ability to model non-local contextual information. These methods exhibit poor compatibility on resource-constrained mobile devices, limiting their deployment in real-time applications. To address this, we propose a Multi-frequency Adaptive Fusion Network (MAFNet), which can produce high-quality disparity maps using only efficient 2D convolutions. Specifically, we design an adaptive frequency-domain filtering attention module that decomposes the full cost volume into high-frequency and low-frequency volumes, performing frequency-aware feature aggregation separately. Subsequently, we introduce a Linformer-based low-rank attention mechanism to adaptively fuse high- and low-frequency information, yielding more robust disparity estimation. Extensive experiments demonstrate that the proposed MAFNet significantly outperforms existing real-time methods on public datasets such as Scene Flow and KITTI 2015, showing a favorable balance between accuracy and real-time performance.

Problem

• Addresses computational overhead in stereo matching cost aggregation
• Enhances non-local contextual modeling in disparity estimation
• Improves real-time performance on resource-constrained mobile devices

Innovation

• Multi-frequency adaptive fusion network for stereo matching
• Adaptive frequency-domain filtering attention module for cost volume
• Linformer-based low-rank attention mechanism for frequency fusion