MORS face recognition datasets

Preprocessed benchmark trees used in Face Recognition via Synthetic Data and Turbocharged Two-Stage Contrastive Fine-Tuning (MORS 2026).

Code & supplementary figures: GitHub — FR-via-SynData-and-TC-2SCFT

Each folder is a complete mirror of the local benchmark tree used in our experiments: training images (originals + synthetics), official test/eval splits, protocol .mat files, and feature caches where present. Nothing is stripped for upload.

Repository layout

tinyface/ — complete tree

• Training_Set/ — per-ID folders with originals + mods 1–5 synthetics
• Testing_Set/ — Gallery_Match/, Gallery_Distractor/, Probe/ images + protocol .mat pair files
• Face_Identification_Evaluation/ — eval protocol .mat files and features/ cache
• features/ — additional gallery/probe/distractor .mat feature blocks
• readme.txt — official TinyFace notes

QMUL/ — complete tree

• training_set/ — per-ID folders with originals + mods 1–5 synthetics
• Face_Identification_Test_Set/ — gallery/, mated_probe/, unmated_probe/ + protocol .mat files
• Face_Verification_Test_Set/ — verification images + positive_pairs_names.mat, negative_pairs_names.mat, features_verification.mat
• Face_Identification_Evaluation/ — eval .mat files, features/, and features_export_baseline/
• Face_Verification_Evaluation/ — verification eval assets
• readme.txt — official SurvFace notes

xqlfw4mors_testsplit/ — paper cohort

• Training_Set/ — 1,318 train IDs / 6,855 images (originals + synthetics)
• Testing_Set/ — 330 held-out eval IDs / 2,185 images
• metadata/ — identity_map_train.csv, identity_map_test.csv, xqlfw_pairs_resolved.csv (623 verification pairs)
• SPLIT_README.md — split definition (80/20, seed 1337, ≥2 originals per train ID)

Synthetics (mods 1–5: LR, distance, HR, POSE3D, passport) live under each benchmark's training layout.

Raw backbone checkpoint

checkpoints/inception_resnet_v1_vggface2_raw.pt — unfine-tuned Inception-ResNet-V1 (vggface2, 512-D). Used for the paper raw baseline and cold-start ft_ogsyn.

import torch
from facenet_pytorch import InceptionResnetV1

ckpt = torch.load("inception_resnet_v1_vggface2_raw.pt", map_location="cpu", weights_only=False)
state = ckpt["state_dict"] if "state_dict" in ckpt else ckpt
model = InceptionResnetV1(pretrained=None)
model.load_state_dict(state, strict=True)
model.eval()

Download

pip install huggingface_hub
huggingface-cli download afro-chai/FR-via-SynData-and-TC-2SCFT --repo-type dataset --local-dir ./mors_data

Or download individual folders from the dataset repo file browser.

Usage with MORS runners

Point runners at the downloaded trees (see GitHub README):

# TinyFace
python model_training/MORS/run_tinyface.py --config experiments/mods_1_5/rebaseline_60_mods15.yaml \
  --tiny_root ./mors_data/tinyface

# QMUL
python model_training/MORS/run_qmul.py --config experiments/mods_1_5/3_ctrl_baseline_60_mods15_qmul.yaml \
  --qmul_root ./mors_data/QMUL

# XQ-LFW paper testsplit
python model_training/MORS/run_xqlfw_testsplit.py --config experiments/xqlfw_testsplit/mods_1_5/rebaseline_60_mods15_testsplit.yaml \
  --xqlfw_root ./mors_data/xqlfw4mors_testsplit

Regenerate locally

Layouts can be rebuilt from official benchmark releases using scripts in the GitHub repo (data/preprocessing_cleaning/, data/IS_pipeline/). The backbone checkpoint can be regenerated with bash data/huggingface/export_raw_backbone.sh.