Sensor and database shift in visible–thermal face recognition

Characterize and mitigate the effects of sensor and database shift on visible–thermal cross-spectral face recognition, with the aim of improving generalization across acquisition conditions and datasets.

Background

The paper evaluates SynThermFace and the PACT adaptation strategy on the Tufts Face Dataset after training on MCXFace. Although PACT improves transfer relative to an untuned EdgeFace model and an xEdgeFace baseline, performance on Tufts remains substantially worse than performance on MCXFace.

The authors interpret this gap as evidence that differences in sensors and databases remain unresolved barriers to reliable cross-database visible–thermal face recognition. Addressing this challenge would improve the robustness and deployment potential of models trained with limited paired data and synthetic visible–thermal supervision.