metadata

task_categories:
  - image-classification
pretty_name: MS-Celeb-1M
size_categories:
  - 1M<n<10M

MS-Celeb-1M (v3)

This is a copy of gaunernst/ms1mv3-wds with gzip compression. Thus, the shards (after decompression) are identical. The compression ratio is around 50%, indicating that the original JPEG images were not compressed much.

This dataset is introduced in the Lightweight Face Recognition Challenge at ICCV 2019. Paper.

There are 5,179,510 images and 93,431 ids. All images are aligned based on facial landmarks predicted by RetinaFace and resized to 112x112.

This was downloaded from https://github.com/deepinsight/insightface/tree/master/recognition/_datasets_ (MS1M-RetinaFace). The original dataset format is MXNet RecordIO. It was converted to WebDataset in this copy here. There are 100 shards in total.

Usage

import webdataset as wds

url = "https://huggingface.co/datasets/gaunernst/ms1mv3-wds-gz/resolve/main/ms1mv3-{{0000..0099}}.tar.gz"
ds = wds.WebDataset(url).decode("pil").to_tuple("jpg", "cls")

img, label = next(iter(ds))