Datasets:

jk-junhokim
/

lfd-dataset-iccp2026

Tags:

lensless-imaging

computational-photography

face-recognition

face-verification

iccp-2026

License:

Dataset card Files Files and versions

xet

Community

Request access to the LFD dataset

Access is reviewed and granted manually by the authors. Requests are typically processed within a few days.

These are identifiable face images of consenting study participants, released
under the LFD Data Use Agreement (see the LICENSE file in this repository).
By requesting access you agree to: (1) use the data for non-commercial
academic research only, (2) not redistribute it in any form, (3) not attempt
to re-identify participants, and (4) cite the LFD paper in any resulting
publication.

LFD Dataset

Real-world lensless face captures accompanying LFD: Enabling Real-World Lensless Face Recognition with a Large-Scale Dataset (ICCP 2026).

Code: https://github.com/jk-junhokim/lfd-face-recog

Distributed on request. Do not redistribute. The images are identifiable faces of consenting study participants; no name→id mapping is included in this release and none should be reconstructed.

Layout

LFD_Final_Dataset/
├── proto1/        104 subjects — phase-mask prototype 1
│   ├── meas/      <id>/meas_*.png     raw lensless measurements
│   ├── recon/     <id>/recon_*.png    FlatNet reconstructions
│   └── webcam/    <id>/webcam_*.png   lensed reference (webcam), same sessions
├── proto2/          7 subjects — phase-mask prototype 2
│   ├── meas/      <id>/meas_*.png
│   └── recon/     <id>/recon_*.png
└── binary_cam/      4 subjects — binary (MLA) mask camera
    ├── meas/      <id>/meas_*.png
    └── recon/     <id>/recon_*.png

Subject IDs are integers (1, 2, …). IDs are per-prototype and do not correspond across proto1 / proto2 / binary_cam.

Each subject has up to 188 captures per modality, spanning indoor and outdoor sessions. Capture index ranges define the evaluation categories (easy, hard, indoor, outdoor, complete); the pair files that encode them ship with the code under data/file_lists/evaluations/.

What is not here

proto2 and binary_cam have no webcam/lensed reference images. Only proto1 was captured with a paired webcam. Evaluations that need a lensed reference (the oracle upper bound, real-to-standard, image quality) are therefore proto1-only.

Detected/cropped face images are not included. Generate them with the released detector checkpoint:

python data/split_generation/generate_annotations.py \
    --detect-checkpoint $LENSLESS_CHECKPOINT_ROOT/<flatrcnn_experiment>/best.tar \
    --psf data/psf/psf_30cm_flatlens-v2_y16.tiff \
    --dataset-root $LENSLESS_DATASET_ROOT/proto1/meas \
    --subjects 1-104 \
    --output data/file_lists/train_lists/proto1_annots.txt

Model checkpoints are distributed separately; see the "Data and Checkpoints" section of the code README.
Raw name-keyed session folders are withheld for privacy.

Quick start

# after `hf download jk-junhokim/lfd-dataset-iccp2026 --repo-type dataset --local-dir ./LFD_Final_Dataset`
export LENSLESS_DATASET_ROOT=/path/to/LFD_Final_Dataset
export LENSLESS_CHECKPOINT_ROOT=/path/to/checkpoints

# lensed (webcam) oracle upper bound — proto1 only
python experiments/scripts/final_tables/eval_oracle_webcam.py \
    --file-root $LENSLESS_DATASET_ROOT/proto1/webcam

License

Released under the LFD Data Use Agreement v1.0 — see LICENSE. Non-commercial academic research only; redistribution is not permitted.

Citation

@inproceedings{kim2026lfd,
  author    = {Kim, Junho and Khan, Salman S. and Wan, Sara and Kuye, Tomi and Veeraraghavan, Ashok},
  title     = {{LFD}: Enabling Real-World Lensless Face Recognition with a Large-Scale Dataset},
  booktitle = {Proceedings of the IEEE International Conference on Computational Photography (ICCP)},
  year      = {2026},
}

Downloads last month: 37

Total file size:

148 GB