X-Palm dataset bridges gap between controlled enrollment and smartphone palmprint auth

Researchers from the University of Toronto and other institutions have released X-Palm, a novel cross-domain dataset designed to tackle the fundamental challenge of palmprint authentication: the gap between controlled enrollment and unconstrained smartphone verification. The dataset comprises 6,006 palm images from 103 individuals (206 hands), captured under two distinct domains: a controlled multispectral scanner developed by the team, and a participant-driven smartphone setup that simultaneously varies hardware, hand pose, illumination, background, camera distance, perspective, and palm surface conditions (moisture, occlusion). This paired-identity structure is unprecedented—each identity appears in both domains, enabling direct domain adaptation research.

Extensive benchmarks using 12 state-of-the-art palmprint recognition models reveal a critical finding: methods that perform well on controlled datasets suffer severe performance collapse when evaluated on X-Palm's unconstrained smartphone images. Conversely, models trained on X-Palm show consistent robustness across domains, positioning the dataset as a valuable resource for real-world generalization. The researchers provide public access to the dataset and benchmarking code, aiming to push the field toward deployable, cross-domain palmprint authentication.