WaveMamba: Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection

TL;DR

WaveMamba introduces a wavelet-based fusion method for RGB-Infrared object detection, significantly improving accuracy by effectively integrating frequency features and reducing information loss.

Contribution

The paper presents a novel wavelet-driven fusion framework with an improved detection head and a specialized fusion block, advancing multi-modality object detection techniques.

Findings

Achieves 4.5% higher mAP on four benchmarks

Effectively fuses low-/high-frequency features via wavelet transforms

Outperforms existing state-of-the-art methods

Abstract

Leveraging the complementary characteristics of visible (RGB) and infrared (IR) imagery offers significant potential for improving object detection. In this paper, we propose WaveMamba, a cross-modality fusion method that efficiently integrates the unique and complementary frequency features of RGB and IR decomposed by Discrete Wavelet Transform (DWT). An improved detection head incorporating the Inverse Discrete Wavelet Transform (IDWT) is also proposed to reduce information loss and produce the final detection results. The core of our approach is the introduction of WaveMamba Fusion Block (WMFB), which facilitates comprehensive fusion across low-/high-frequency sub-bands. Within WMFB, the Low-frequency Mamba Fusion Block (LMFB), built upon the Mamba framework, first performs initial low-frequency feature fusion with channel swapping, followed by deep fusion with an advanced gated attention mechanism for enhanced integration. High-frequency features are enhanced using a strategy that applies an ``absolute maximum" fusion approach. These advancements lead to significant performance gains, with our method surpassing state-of-the-art approaches and achieving average mAP improvements of 4.5% on four benchmarks.