OpenVeinNet, a model for finger-vein verification, reported the lowest equal error rate, or EER, in four of five full-subject splits. In the fifth, it remained competitive. For the BCDE → A split, its EER was 3.17%, compared with 6.20% for VeinAttNet. In these comparisons, a lower EER represents a better verification result.
The test was designed around an open-set problem: the system had to handle identities that were unseen during training and reject non-enrolled probes. In the cross-dataset protocol, four datasets supplied training data and the fifth was held out, so evaluation identities remained unseen during training.
What OpenVeinNet puts together
The benchmark used five public finger-vein image datasets. FV-300 was reported with 300 classes and 27,600 images; MMCBNU with 600 classes and 6,000 images; FV-USM with 492 classes and 5,904 images; PolyU with 218 classes and 872 images; and VERA with 220 classes and 440 images. The paper also says 124 low-quality FV-300 images were removed.
OpenVeinNet combines Dynamic Snake Convolution for local curvilinear and tubular features with graph convolution for long-range vascular relationships. A Centroid Angular Hybrid Loss is used to produce compact, angularly separated embeddings, and verification scores are generated with cosine similarity.
Strong scores on held-out identities
In one half-subject unknown-rejection split, labeled ABDE → C, the model produced an area-under-curve score of 97.72% and an EER of 7.48%. Its true-accept rate was 53.52% at a 1% false-accept rate and 95.47% at 10%. On ABCE → D, it recorded the best reported AUC, 98.67%, and the lowest EER, 5.34%.
The statistical comparison used two-sided exact sign-flip permutation tests with Holm correction, alongside paired two-sided Wilcoxon signed-rank tests, on 20 matched runs for each comparator. Every listed comparison was significant after correction. Mean EER differences favored OpenVeinNet by 15.17 percentage points against ArcVein, 13.05 against LGFIN, 13.66 against FV-ViT, and 4.44 against VeinAttNet.
The component tests
Component ablations gave the full model a mean EER of 7.44%, compared with 10.96% for a DSConv stem without graph modelling, 11.64% for a standard stem with graph modelling, and 13.59% for a standard stem without graph modelling. The complete configuration was the lowest of the four reported variants.
A separate loss comparison gave CAH an EER of 7.44%, versus 26.68% for mean squared error and 33.98% for negative log likelihood. Against angular-margin alternatives, CAH's mean EER was 7.44%, compared with 7.83% for ArcFace, 8.20% for MagFace and 8.22% for AdaFace.
The architecture tests also examined kernel sizes and graph depth. A progressively decreasing Dynamic Snake Convolution sequence of 9, 7 and 3 had the best reported kernel-ablation EER, 10.82%. Four Grapher Blocks in the first stage and six in the second produced the best graph-depth ablation EER, 10.595%.
A caveat at the strictest threshold
In the low-sample VERA evaluation, the true-accept rate at a 0.1% false-accept rate was zero for every method. The paper cautions that the few genuine comparisons created a coarse score distribution and limited resolution at very low false-accept thresholds.
The cost of running the model
Supplementary intra-database open-set tests reported the lowest EER for OpenVeinNet on each of three datasets: 0.39% on FV-300, 7.06% on FV-USM and 4.07% on MMCBNU.
On FV-300 with a batch size of one, the proposed method used 4.76 million parameters and 5.07 billion FLOPs. Its reported latency was 55.82 milliseconds on a CPU and 13.00 milliseconds on a GPU, with throughput of 17.92 images per second on a CPU and 76.94 on a GPU.
The work is still a preprint
The document is an arXiv version 1 preprint dated 26 Aug 2026.
Paper data and sources
Original title: OpenVeinNet: Robust Open-Set Finger Vein Verification with Dynamic Snake Convolution and Graph Learning
Authors: Sushrut Patwardhan, Raghavendra Ramachandra
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text