IMPOSE: Identity-Consistent Multi-Pose Generation of Contactless Fingerprints

Identity-Consistent Multi-Pose Generation of Contactless Fingerprints

Zhiyu Pan, Xiongjun Guan, Jianjiang Feng, and Jie Zhou

Department of Automation, Tsinghua University

Under review. Preprint available at arXiv:2605.03830.

---

Generated rolled fingerprint identity and its contactless multi-pose renders under three roll angles (−35°, 0°, +27°).

---

🔍 Overview

IMPOSE is a physics-inspired generative framework for synthesizing identity-consistent, multi-pose contactless fingerprint samples to improve cross-modal fingerprint recognition. This repository provides the full code, pre-trained model weights, and large-scale generated datasets.

The framework integrates three core stages:

• 🧬 Identity Generation: Synthesizing master rolled fingerprints via latent diffusion with discrete VQ-VAE codebook.
• 🌓 Cross-Modal Translation: Translating rolled to contactless modality via ControlNet, guided by Sauvola binarization as a deterministic identity anchor.
• 🔄 Multi-Pose Simulation: Projecting contactless textures onto 3D finger models and rendering under diverse roll angles.

---

📦 Released Data

We publicly release IMPOSE-SYN, the large-scale synthetic dataset generated by IMPOSE, usable for training and fine-tuning fixed-length fingerprint descriptors.

Dataset	Description	Size	Download
IMPOSE-SYN	Rolled fingerprints (gallery) + multi-pose contactless fingerprints (query)	1,712 + 15,136 images	Google Drive

All images are normalized to 512 × 512 px at 500 PPI, aligned with the standard fingerprint coordinate system. Identities are sequentially numbered 00000–01711.

---

🛠️ Environment Setup

• GPU with CUDA support
• Python ≥ 3.8

Install dependencies:

pip install numpy omegaconf Pillow opencv-python einops pytorch_lightning
pip install scipy matplotlib scikit-learn scikit-image tqdm

The code has been tested with PyTorch 2.0.0 + CUDA 11.8:

pip install torch==2.0.0 torchvision==0.15.1 --index-url https://download.pytorch.org/whl/cu118

This project builds on Latent Diffusion Models and taming-transformers. Relevant modules are already included under ldm/.

---

📥 Pre-Trained Model Weights

Download the checkpoints and place them under models/:

Stage	Description	Download	Target Directory
Rolled-Gen	Rolled fingerprint identity synthesis	Google Drive	models/fingerprint_ldm_rolled_512/
Cross-Modal	Contactless texture & modality transfer	Google Drive	models/fingerprint_c2cl_512/

Expected structure:

models/
├── fingerprint_ldm_rolled_512/
│   └── model_ldm_rolled.ckpt
└── fingerprint_c2cl_512/
    └── model_unsw_flat.ckpt

---

🚀 Quick Start

Example inputs and outputs are provided under example/ for a quick trial.

1. Generate Rolled Fingerprints

python Rolled_FP_generation.py \
  --n_samples 4 --ddim_steps 50 --ddim_eta 0.0 --nums 10

Argument	Default	Description
--nums	10000	Total number of images to generate
--ddim_steps	50	Number of DDIM sampling steps
--ddim_eta	0.0	DDIM stochasticity (0 = deterministic)
--outdir	./example	Output directory

2. Quality Filtering (Optional)

Filter samples with NFIQ 2.0 > 0.55 and foreground ratio > 60%. Pre-filtered examples: example/source_rolled/.

3. Ridge Enhancement (Sauvola)

python sauvola_simulation_contact.py

Reads from example/source_rolled/ and example/mask/, writes to example/binary_enhancement/.

4. Cross-Modal Texture Generation

python cross_modal_generation.py \
  --img-folder example/binary_enhancement --ddim_steps 50

Outputs to example/contactless_texture/.

5. Multi-Pose Texture Mapping

python multi_pose_CL_generation.py

Uses 3D point clouds (example/finger3d/) and pose parameters (example/finger3d_pose/). Outputs to example/multi_pose_CL/.

---

📂 Repository Structure

IMPOSE/
├── Rolled_FP_generation.py          # Stage 1: rolled fingerprint generation
├── sauvola_simulation_contact.py    # Stage 2a: Sauvola binarization
├── cross_modal_generation.py        # Stage 2b: cross-modal texture translation
├── multi_pose_CL_generation.py      # Stage 3: physics-based multi-pose simulation
├── configs/
│   ├── fingerprint-ldm-vq-128-512-rolled.yaml
│   └── fingerprint-cldm-vq-128-512-contactless-eval.yaml
├── ldm/                             # Latent Diffusion Model modules
├── mapping/                         # 3D pose correction, unfolding, projection
├── example/                         # Example inputs & outputs for all stages
├── models/                          # (Download) Pre-trained checkpoints
└── README.md

---

📄 Citation

@article{pan2026impose,
  title={Identity-Consistent Multi-Pose Generation of Contactless Fingerprints},
  author={Pan, Zhiyu and Guan, Xiongjun and Feng, Jianjiang and Zhou, Jie},
  journal={arXiv preprint arXiv:2605.03830},
  year={2026}
}

---

⚠️ License

This project is licensed under the Apache License 2.0. See LICENSE for details.

---

📬 Contact

For technical issues: Zhiyu Pan — pzy20@mails.tsinghua.edu.cn

For research collaboration: Jianjiang Feng — jfeng@tsinghua.edu.cn