The cheat sheet for teaching cameras to see in the dark

What it does

This is an “awesome list” in the classic mold: a hand-curated catalog of deep learning methods that fuse RGB and thermal (infrared) imagery. It organizes the field into six tasks — multispectral pedestrian detection, aerial object detection, semantic segmentation, salient object detection, crowd counting, and fusion tracking — and for each, links papers, code repos, and datasets. Think of it as a literature review you can actually click through.

The interesting bit

The dataset tables are the quiet star. For pedestrian detection alone it catalogs fourteen datasets (KAIST, LLVIP, M3FD, MMPD, and friends) with years, modalities, image counts, and class counts — the kind of comparison shopping that normally takes an afternoon of tab-hopping. It also collects the various “fixed” KAIST annotation sets, which anyone who has worked with that dataset will appreciate deeply.

Key highlights

• Covers six RGB-T task areas with linked papers, code, and datasets
• Includes foundation-model-era work like M-SpecGene (ICCV 2025) and UniRGB-IR, not just older CNN-era fusion papers
• Dataset tables with modality, size, and class breakdowns for quick benchmark selection
• Curated KAIST annotation fixes and evaluation tooling
• Actively maintained, with updates noted through recent conference cycles

Caveats

• It’s a link collection, not a library — no code runs here, and some entries are PDF-only with no released implementation
• Coverage depth varies by task; pedestrian detection is far more fleshed out than, say, crowd counting

Verdict

If you’re starting RGB-thermal fusion research or picking a benchmark, bookmark it — it saves real time. If you wanted runnable tooling or a survey with commentary, this is a map, not the territory.