AGVBench evaluates 30 data augmentation strategies across five public palm- and finger-vein datasets and seven backbone architectures, including CNNs, vision transformers, and vein-specific models. The authors report that multi-image mixing methods such as MixUp, PuzzleMix, and StarMixup generally achieve the strongest clean recognition results, but often have poor calibration and greater vulnerability to adversarial perturbations. Severe geometric transformations frequently reduce performance, possibly because of feature misalignment or spatial cropping. Results also vary between palm- and finger-vein datasets, motivating reliability-oriented evaluation beyond accuracy.
No heat snapshots are available in the last 24 hours.