A new paper proposes that most thermal-visible fusion detectors are doing expensive work on pixels that don't need it.
Researchers behind the arXiv preprint noticed that the majority of any image — sky, ground, empty road — is smooth background that a simple single-modality model handles fine. Their proposed system, built for RGB-T (visible plus thermal infrared) detection, runs a lightweight first pass to flag candidate regions, then applies the costly cross-modality fusion only to those sparse proposals. The result is a two-stage pipeline: a fast, modality-specific sweep for high recall, followed by a fusion-driven refinement pass that kills false positives and tightens bounding boxes. The team reports competitive detection accuracy with substantially fewer parameters and lower compute, and says the approach scales to high-resolution images.
RGB-T detection matters most in conditions where cameras alone fail — night driving, fog, search and rescue — so efficiency gains here have real deployment value, not just benchmark value. Most current approaches run dual backbones and fuse features across the entire image, a design that looks increasingly hard to justify if targeted fusion can match it.
The paper's results are described as "competitive" with full-fusion methods, not clearly superior — so the question of whether selective fusion is genuinely sufficient, or merely close enough for some use cases, remains open.