Visible-thermal face recognition matches thermal probe images against visible enrollment templates, a technique useful in low-light or surveillance settings. According to the paper, existing benchmarks for this task are limited by small gallery sizes, typically around 100 subjects, which restricts the reliability of performance comparisons.

The new VTF900 dataset increases that scale to 900 subjects, offering a larger and more realistic evaluation resource. The abstract presents this as a direct response to the prior limitation, though it does not detail the collection process or specific results.

As a benchmark contribution, VTF900 provides a common ground for testing algorithms under a more demanding gallery size. The paper's main significance lies in this expansion, which may lead to more robust assessments of visible-thermal face recognition systems.