CALFW HF-ready

This folder packages the local CALFW (Cross-Age LFW) images as a Hugging Face imagefolder dataset with two splits and a verification-pairs CSV.

Splits

• aligned: 112x112 ArcFace-aligned faces (from data/evaluation/CALFW/aligned/).
• raw: original 250x250 LFW-style faces (from data/evaluation/CALFW/raw/).

Both splits contain the same 12174 images and the same identity labels.

Layout

calfw/
├── README.md
├── pairs.csv
├── aligned/
│   ├── images/<shard>/<file>.jpg
│   └── metadata.csv
└── raw/
    ├── images/<shard>/<file>.jpg
    └── metadata.csv

metadata.csv columns

• file_name: relative image path used by ImageFolder, e.g. images/000/AJ_Cook_0001.jpg.
• label: numeric identity label, shared across the two splits.
• label_name / identity: identity name corresponding to label.
• image_num: per-identity image index from the original filename (e.g. 0001).
• variant: aligned or raw.
• source_filename: original CALFW filename.

pairs.csv columns

pairs.csv mirrors txts/pairs_CALFW.txt (6000 verification pairs; 3000 same-identity cross-age positives and 3000 negatives, split into 10 folds of 600 pairs each).

• pair_id (0..5999), fold_id (1..10), fold_position (0..299).
• is_same: 1 for positive pairs, 0 for negatives.
• image_a, image_b: bare filenames as in the upstream pairs file.
• image_a_aligned_path, image_b_aligned_path: path under the aligned split.
• image_a_raw_path, image_b_raw_path: path under the raw split.

Local Stats

• Images per split: 12174
• Unique identities: 4025
• Identities with one image: 1
• Verification pairs: 6000 (3000 positive / 3000 negative)
• Folds: 10 x 600 pairs

Loading

from datasets import load_dataset

ds = load_dataset("imagefolder", data_dir="data/evaluation/huggingface/calfw")
ds["aligned"], ds["raw"]

import pandas as pd

pairs = pd.read_csv("data/evaluation/huggingface/calfw/pairs.csv")

Notes

CALFW (Cross-Age LFW) is described by its authors as a verification benchmark that emphasizes age variation between the two faces in each positive pair. Check the original dataset terms before publishing or redistributing it.