Research map / 3D vision / Depth and stereo

Towards Explainable and Unprecedented Accuracy in Matching Challenging Finger Crease Patterns

CVPR 2025 · 3D vision › Depth and stereo · approach: Camera model and calibration for SfM · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a keypoint-correspondence graph neural network (CGN) for matching deformed finger knuckle crease patterns, releases a 805,768-image multi-pose knuckle video dataset, and gives a first uniqueness estimate for 2D knuckle patterns.

Problem

Matching contactless finger knuckle images with severe pose deformations is very inaccurate, and existing methods are not interpretable for forensic use.

Approach

Detects knuckle crease keypoints with a SuperPoint-style CNN, establishes correspondences via attentional GNN with Sinkhorn, builds k-NN graphs over matched keypoints, and scores pairs using node-dimension graph convolution plus cross-node cosine similarity tracked across layers; also estimates uniqueness with a Gaussian mixture model and Poisson false-random-correspondence model.

Key results

• On the new 351-subject multi-pose video dataset, the method reaches 89.99% GAR at FAR=10^-5 versus 67.22% for the best baseline (EfficientNetv2-m).
• On the most challenging Finger Knuckle v3.0 dataset, it achieves 55.41% GAR at FAR=10^-5 versus 1.04% for the best baseline (FKNet).
• Cross-database on the Hand Dorsal dataset, it reaches 80.19% GAR at FAR=10^-5 versus 53.11% for EfficientNetv2-m.

Measured effects reported in the paper

Surprising findings

• Standard CNN and ViT baselines reach 0.00% GAR at FAR=10^-5 on the Finger Knuckle v3.0 dataset, showing the difficulty of cross-pose matching.

Limitations

• GNN architecture lacks spatial awareness, which the authors identify as a gap.
• Uniqueness numbers depend on the low image resolution used and are far below fingerprint-level resolution (about 500 dpi).
• Results on the most challenging dataset use only 104 subjects in the second session.

Experimental setup

Datasets: Multi-pose finger knuckle video dataset (new) Hand Dorsal Image Dataset [47] Finger Knuckle v3.0 [11]

Benchmarks: Multi-pose finger knuckle video dataset (new) Hand Dorsal Image Dataset [47] Finger Knuckle v3.0 [11]

Metrics: GAR at FAR 10^-5 GAR at FAR 10^-4 GAR at FAR 10^-2 ROC curves

Baselines: ResNet-101 DenseNet-161 EfficientNetv2-m ViT-B FKNet RFNet-RSIL STResNet GMN SimGMN MGMN ERIC

Builds on: SuperPoint SuperGlue GMN SimGNN ERIC

Closest papers by meaning

• Diff-Palm: Realistic Palmprint Generation with Polynomial Creases and Intra-Class Variation Controllable Diffusion Models (CVPR 2025 · similarity 0.81)
• PCE-Palm: Palm Crease Energy Based Two-Stage Realistic Pseudo-Palmprint Generation (AAAI 2024 · similarity 0.80)
• Learning Dense Hand Contact Estimation from Imbalanced Data (NeurIPS 2025 · similarity 0.80)
• EasyHOI: Unleashing the Power of Large Models for Reconstructing Hand-Object Interactions in the Wild (CVPR 2025 · similarity 0.79)
• DICE: End-to-end Deformation Capture of Hand-Face Interactions from a Single Image (ICLR 2025 · similarity 0.79)
• HOLD: Category-agnostic 3D Reconstruction of Interacting Hands and Objects from Video (CVPR 2024 · similarity 0.79)