LamTNguyen/stage2-exp0-ckpt50-synth500x10-sd302a-20260609

Stage2 Exp0 ckpt50 synthetic SD302A 500x10

This dataset contains synthetic rolled fingerprint images generated with the Stage-2 Exp0 ControlNet checkpoint at step 50k. The images are arranged in an SD302A-compatible folder structure:

images/challengers/<SENSOR>/roll/png/<ID>_<SENSOR>_roll_<IMPRESSION>.png

Contents

• —images/challengers/: 5,000 synthetic PNG images.
• —manifest.json: dataset counts by sensor.
• —selection/selection.csv: source real condition image selected for each synthetic sample.
• —selection/selection_summary.json: label quotas and eligible-label counts.
• —resnet50_real_sd302a_val_eval/: classifier and verification evaluation outputs.

Generation Setup

• —Generator: Stage-2 Exp0 self-condition ControlNet.
• —Checkpoint: core_exp0_self_condition_unetinit_bs256_50k/step_0050000.
• —Base UNet: official IMPOSE Stage-1 UNet.
• —DDIM steps: 50.
• —Control scale: 1.0.
• —Conditions: Sauvola ridge/control maps from SD302A roll images.
• —Selection seed: 20260609.

Dataset Size

• —Labels: 500 fingers.
• —Impressions per label: 10.
• —Images: 5,000.

Sensor label counts:

Sensor H has fewer labels because only 21 SD302A H labels have at least 10 roll impressions.

Evaluation Snapshot

A ResNet50 ArcFace/CosFace classifier was trained on this synthetic dataset with the same paperlike setting used for the SD302A real-data classifier:

• —image size: 224
• —embedding dim: 512
• —epochs: 200
• —batch size: 128
• —loss: CosFace, s=64, m=0.35
• —augmentation: enabled
• —identity mode: finger

Real SD302A held-out validation split, filtered to these 500 labels:

• —Images: 1,000 real images.
• —Top1: 0.601.
• —Top5: 0.808.
• —Verification EER: 0.088.
• —TAR@FAR 1%: 0.496.
• —TAR@FAR 0.1%: 0.180.

For comparison, the real-data-trained classifier evaluated on the same 500 labels gives EER 0.122, TAR@FAR 1% 0.488, TAR@FAR 0.1% 0.198.

Notes

This dataset is intended for internal research experiments on synthetic fingerprint data utility. Check the applicable NIST SD302A terms and local biometrics data handling requirements before redistribution or downstream use.