Research map / Image generation / Autoregressive and latent generation
PVTree: Realistic and Controllable Palm Vein Generation for Recognition Tasks
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 PVTree, a framework that builds identity-defining 3D palm vascular trees via improved CCO, renders multi-view 2D projections, and translates them to realistic palm vein images with PCE-Palm to generate synthetic palm vein training data.
Problem
Palm vein recognition lacks large public datasets, and existing synthetic palm vein generators produce unrealistic patterns or cannot control identity and intra-class variation.
Approach
Define identity by a 3D palm vascular tree generated with an improved Constrained Constructive Optimization algorithm guided by anatomical trunks and a trajectory simulation mechanism; produce same-identity variation by multi-view projection, augmentation, and l2-constrained noise sampling in a PCE-Palm image-to-image translator.
Key results
- Recognition model trained on 4000-ID PVTree synthetic data reaches average TAR@1e-4 0.9448 vs 0.9380 for real data; with 3000 IDs already exceeds real data (0.9398).
- Mixing real data with 4000-ID synthetic data gives average TAR@1e-4 0.9713 and EER 0.0060.
- FID of 37.08 versus 56.17 (Syn-sPVDB) and 50.37 (NS-PVDB) on synthetic quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Generate training data with PVTree, 4000 IDs rather than real data | ResNet101 palm-vein recognition · Average across CASIA, PolyU, HFUT, and TongJi | TAR@1e-4: 0.9380 -> 0.9448 | Helped |
| Mix real data with PVTree synthetic data | ResNet101 palm-vein recognition · Average across CASIA, PolyU, HFUT, and TongJi | TAR@1e-4: 0.9448 -> 0.9713 | Helped |
| Use PVTree-generated images instead of PCE-Palm-generated images for recognition training | ResNet101 palm-vein recognition · Average across CASIA, PolyU, HFUT, and TongJi | TAR@1e-4: 0.8383 -> 0.9448 | Helped |
| Set the noise-input l2 distance to 0.5 rather than 0.3 (ablation) | PVTree intra-class variation controls · Average across CASIA, PolyU, HFUT, and TongJi | TAR@1e-4: 0.9057 -> 0.9239 | Helped |
| Set the noise-input l2 distance to 0.5 rather than 0.7 (ablation) | PVTree intra-class variation controls · Average across CASIA, PolyU, HFUT, and TongJi | TAR@1e-4: 0.8956 -> 0.9239 | Helped |
| Add image enhancement disturbances to simulate palm posture (ablation) | PVTree intra-class variation controls · Average across CASIA, PolyU, HFUT, and TongJi | TAR@1e-4: 0.9239 -> 0.9298 | Helped |
| Add multi-view projection disturbances (ablation) | PVTree intra-class variation controls with image enhancement · Average across CASIA, PolyU, HFUT, and TongJi | TAR@1e-4: 0.9298 -> 0.9448 | Helped |
Surprising findings
- Synthetic-only training with 3000-4000 identities surpasses real-data training on average TAR@1e-4 in the reported experiments.
Limitations
- Comparisons to Syn-sPVDB and NS-PVDB rely on reported samples because their code and data were not released, limiting fair comparison.
- Synthetic data only matches real-data performance at larger identity counts (3000+ IDs), and the advantage over real data is small (0.9448 vs 0.9380).
Experimental setup
Datasets: CASIA PolyU TongJi PUT HFUT
Benchmarks: CASIA PolyU HFUT TongJi
Metrics: EER TAR@FAR=1e-4 FID Wang17 vein quality score GLCM feature overlap
Baselines: NS-PVDB Syn-sPVDB PCE-Palm IDiff-Face Real data
Builds on: Constrained Constructive Optimization PCE-Palm ArcFace Kamiya optimization
Closest papers by meaning
- Diff-Palm: Realistic Palmprint Generation with Polynomial Creases and Intra-Class Variation Controllable Diffusion Models (CVPR 2025 · similarity 0.84)
- PCE-Palm: Palm Crease Energy Based Two-Stage Realistic Pseudo-Palmprint Generation (AAAI 2024 · similarity 0.84)
- vesselFM: A Foundation Model for Universal 3D Blood Vessel Segmentation (CVPR 2025 · similarity 0.78)
- Hand1000: Generating Realistic Hands from Text with Only 1,000 Images (AAAI 2025 · similarity 0.77)
- Analyzing the Synthetic-to-Real Domain Gap in 3D Hand Pose Estimation (CVPR 2025 · similarity 0.77)
- Towards Explainable and Unprecedented Accuracy in Matching Challenging Finger Crease Patterns (CVPR 2025 · similarity 0.76)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.