FlowPalm: Optical Flow Driven Non-Rigid Deformation for Geometrically Diverse Palmprint Generation

Yuchen Zou; Huikai Shao; Lihuang Fang; Zhipeng Xiong; Dexing Zhong

cs.CVApr 11, 2026

FlowPalm: Optical Flow Driven Non-Rigid Deformation for Geometrically Diverse Palmprint Generation

Authors: Yuchen Zou, Huikai Shao, Lihuang Fang, Zhipeng Xiong, Dexing Zhong

Organizations: Xi’an Jiaotong University · Sichuan Digital Economy Industry Development Research Institute · Southern University of Science and Technology

Abstract

Recently, synthetic palmprints have been increasingly used as substitutes for real data to train recognition models. To be effective, such synthetic data must reflect the diversity of real palmprints, including both style variation and geometric variation. However, existing palmprint generation methods mainly focus on style translation, while geometric variation is either ignored or approximated by simple handcrafted augmentations. In this work, we propose FlowPalm, an optical-flow-driven palmprint generation framework capable of simulating the complex non-rigid deformations observed in real palms. Specifically, FlowPalm estimates optical flows between real palmprint pairs to capture the statistical patterns of geometric deformations. Building on these priors, we design a progressive sampling process that gradually introduces the geometric deformations during diffusion while maintaining identity consistency. Extensive experiments on six benchmark datasets demonstrate that FlowPalm significantly outperforms state-of-the-art palmprint generation approaches in downstream recognition tasks. Project page: https://yuchenzou.github.io/FlowPalm/

Figures & tables

Panel of Figure 1 — Figure 1: Overview of the proposed method and its performance comparison with existing methods. (a) We introduce optical flow driven non-rigid deformation modeling to overcome the limitations of handcrafted simple transformations, enabling the synthesis of geometrically diverse and identity-consistent palmprints. (b) Experimental results on six public databases demonstrate that FlowPalm achieves superior recognition performance.

Figure 2: Overview of the deformation prior construction. Deformation field is estimated between real palmprint pairs and filtered through smoothness and identity checks to build a reliable deformation library for generation.

Figure 3: Statistical distribution and visualization of deformation quality assessment.

Figure 4: Overview of the proposed deformation-driven three-stage generation strategy. A deformation field sampled from the deformation library guides the synthesis process: Stage I generates deformed and appearance-diverse principal lines under warped crease conditions and random noise; Stage II produces identity-consistent deformed textures under warped crease conditions and warped crease noise; and Stage III refines texture details through unconditional denoising.

Figure 5: Visualization of the intermediate results of our deformation-driven denoising process.

Setting	Method	Venue	Geo Trans.	Verification Datasets						Avg
Setting	Method	Venue	Geo Trans.	XJTU	MPD	TCD	CASIA	IITD	PolyU	Avg
w/o Aug.	RPG-Palm [ 33 ]	ICCV’23	Line Osc.	0.01	0.22	0.24	2.60	0.46	0.44	0.66
	PCE-Palm [ 12 ]	AAAI’24	Affine	11.38	27.44	34.09	24.73	69.39	33.39	33.40
	UAA [ 14 ]	ICCV’25	Affine	14.04	49.55	79.09	55.47	33.75	76.22	51.35
	Diff-Palm [ 13 ]	CVPR’25	$\times$	2.52	3.52	16.08	4.80	7.58	24.53	9.84
	PFIG-Palm [ 50 ]	TIP’25	Line Osc.	8.84	21.50	58.50	4.80	61.45	66.96	37.01

Table 1: Performance analysis of TAR@FAR=

10^{-6}

(%) on six benchmark datasets with and without data augmentation. Geo Trans. denotes the geometric transformation type adopted in each generation method. Best and second-best results are highlighted.

Setting	Method	Venue	Geo Trans.	Verification Datasets						Avg
Setting	Method	Venue	Geo Trans.	XJTU	MPD	TCD	CASIA	IITD	PolyU	Avg
Real Only	-	-	-	78.05	73.02	88.70	32.60	73.54	95.63	73.59
Syn. Only	Diff-Palm [ 13 ]	CVPR’25	$\times$	72.25	63.09	93.51	66.40	84.55	94.30	79.02
	PFIG-Palm [ 50 ]	TIP’25	Line Osc.	43.06	40.67	82.20	47.80	74.45	83.79	62.00
	FlowPalm (Ours)	-	Non-rigid	86.33	80.37	97.37	66.53	87.32	93.29	85.20
Syn. $+$ Real	Diff-Palm [ 13 ]	CVPR’25	$\times$	90.05	84.41	97.42	77.87	95.30	98.36	90.57

Table 2: Comparison of training strategies. Syn. + Real : mixed training; Syn.

\to

Real : pre-training to fine-tuning.

Method	Configs				Score Distributions			TAR@FAR= $10^{-6}\uparrow$
Method	Deform Sel.	Crease Warp	Noise Warp	Drop Time $\tau_{u}$	Fréchet Distance $\downarrow$	Inter-cls Distance $\uparrow$	Intra-cls Distance $\downarrow$	w/o Aug.	w/ Aug.
RPG-Palm [ 33 ]	-	-	-	-	0.3035	0.9265	0.3341	0.66	63.73
PCE-Palm [ 12 ]	-	-	-	-	0.3738	0.8439	0.5657	33.40	40.98
Diff-Palm [ 13 ]	-	-	-	-	0.2838	0.9181	0.2592	9.84	79.02
PFIG-Palm [ 50 ]	-	-	-	-	0.2372	0.9425	0.3525	37.01	62.00
FlowPalm (Ours)	✓	✓	✓	0.25	0.1503	0.9559	0.3766	58.95	85.20

Table 3: Score distribution comparison and ablation study of the proposed FlowPalm. The top section compares FlowPalm with state-of-the-art generative methods. The middle section presents a component-wise ablation on deformation selection (Deform Sel.), crease warping (Crease Warp), and noise warping (Noise Warp). The bottom section analyzes the effect of condition drop start time

\tau_{u}

in texture refinement stage.

Figure 6: Visualization of the effects of crease and noise warping. Compared with the non-warped samples, warping only the crease leaves fine textures unchanged, while warping only the noise keeps the principal lines fixed. Simultaneously warping both preserves identity-consistent details that follow the deformed geometry.

Figure 7: Visual effects of the condition drop start time $\tau_{u}$ . Each pair of rows shows images corresponding to the same generated identity. Insets highlight local texture details as $\tau_{u}$ increases.

Explore similar work

FVeinSyn: Synthetic Finger Vein Image Generator
Aug 27, 2026Yifan Wang, Jie Gui, Adams Wai Kin Kong +6Synthetic Data Generation Synthetic Data Augmentation
School of Cyber Science and Engineering, Southeast University, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision and Management (Southeast University), Ministry of Education; Purple Mountain Laboratories, Nanjing 210000, China · School of Computer Science and Engineering, Nanyang Technological University, Singapore 639798 +5
CAAP: Capture-Aware Adversarial Patch Attacks on Palmprint Recognition Models
Apr 8, 2026Renyang Liu, Jiale Li, Jie Zhang +6Adversarial Patch Attacks Physical Adversarial Attacks
Institute of Data Science, National University of Singapore, Singapore 117602, Singapore · Centre for Frontier AI Research (CFAR), A*STAR, Singapore, 138634 · School of Cyber Science and Engineering, Wuhan University, China, 430072 +2
Identity-Consistent Multi-Pose Generation of Contactless Fingerprints
May 5, 2026Zhiyu Pan, Xiongjun Guan, Jianjiang Feng +1Synthetic Data Augmentation Biometric Identification
Department of Automation, Tsinghua University, Beijing 100084, China

Aug 27, 2026J/K move · Enter open · S save

FVeinSyn: Synthetic Finger Vein Image Generator

Yifan Wang, Jie Gui, Adams Wai Kin Kong, Baosheng Yu+5

School of Cyber Science and Engineering, Southeast University, Nanjing 210000, China · Engineering Research Center of Blockchain Application, Supervision and Management (Southeast University), Ministry of Education; Purple Mountain Laboratories, Nanjing 210000, China · School of Computer Science and Engineering, Nanyang Technological University, Singapore 639798+5

Synthetic Data Generation Synthetic Data Augmentation Image Generation Biometric Identification

Open paper AskPDF arXiv

A major challenge in finger vein recognition is the lack of large-scale public datasets. Existing datasets contain few identities and limited samples per finger, restricting the advancement of deep learning-based methods. To address this, we propose FVeinSyn, a large-scale controllable synthetic data generation framework for finger vein. It explicitly decouples synthesis of vascular topology and imaging appearance to mitigate the limitations caused by insufficient training samples, such as inadequate identity diversity and restricted realism. Specifically: first, a finger vein identity generator models vascular topology under physiological and geometric constraints using stochastic L-systems, producing anatomically valid and identity-distinctive vascular patterns. Then, a cascaded region-aware GAN renders the topological maps into realistic near-infrared images. Finally, an intra-class diversity generator introduces geometric and optical perturbations to simulate realistic intra-class variations. Using FVeinSyn, we generated 500,000 images (10,000 vein identities, 50 samples per identity) and conducted extensive evaluations. Results show that FVeinSyn holds significant advantages in realism, identity diversity, vascular pattern consistency, and intra-class diversity. Models trained with FVeinSyn outperform real-data-only baselines a cross eight public datasets, achieving an average accuracy improvement of 27.43%. The code is available at: https://github.com/EvanWang98/Synthetic-Finger-Vein-Generator.

Abstract

Explore similar work

FVeinSyn: Synthetic Finger Vein Image Generator

CAAP: Capture-Aware Adversarial Patch Attacks on Palmprint Recognition Models

Identity-Consistent Multi-Pose Generation of Contactless Fingerprints