Abstract

This paper studies learned local descriptors for thermal image matching and registration, with emphasis on patch size. We compare a CNN-based and adapted Vision (ViT) Transformer descriptor across patch sizes from 8 × 8 to 96 × 96. Using a fixed-correspondence benchmark, ViT achieves stronger descriptor separability than the CNN, with 16 × 16
producing the lowest FPR@95TPR. A Ground Truth (GT)-labelled matching test shows that this improvement transfers to stricter match correctness, rather than simply increasing match count. In a shared-detector registration pipeline, both learned descriptors substantially outperform RIFT2, with 16 × 16 ViT giving the best geometric accuracy.