biglam/dating-historical-color-images

Dating Historical Color Images

1,326 colour photographs labelled with the decade in which they were taken, 1930s to 1970s. The collection was assembled for Frank Palermo, James Hays and Alexei A. Efros, Dating Historical Color Images (ECCV 2012), which introduced decade estimation from colour photographs as a task — project page.

Where the images came from

From the paper:

Beginning with a collection of approximately 230,000 Flickr images taken prior to 1980, we perform automated removal of monochromatic images. The remaining images are manually inspected to remove non-photographic content (e.g. scans of vintage artwork) and any remaining monochromatic images. Finally, a random subsampling and decimation is performed to create a dataset containing an equal number of historical color images for each decade (1,375 images total).

The decade labels come from Flickr metadata: the date the uploader recorded, not a verified capture date. Uploaders include individuals and cultural heritage institutions. Treat the labels as silver, not gold.

The paper states 1,375 images; this copy holds 1,326. The source does not document the difference.

Contents

decade	images
1930s	265
1940s	266
1950s	265
1960s	265
1970s	265

The near-balance is a product of the sampling, not of what survives.

One train split. The paper's own train/test division is not reproduced here, so accuracy figures obtained on this data are not comparable with the published ones.

Each row is an RGB JPEG (image) plus the decade label. The images are small — across a 78-image sample the longest side never exceeded 500 px. No EXIF, photographer, or Flickr URL is carried, so an image cannot be traced back to its original or deduplicated against Flickr.

Uses and limits

Decade classification, or ordinal regression — the labels are ordered, and treating them as unordered classes throws that away. Also a small, awkward benchmark for temporal estimation.

The paper's argument is that the discriminative signal lives in the evolution of colour imaging processes. A model trained here is therefore learning film stock and scanning characteristics as much as depicted content, and should not be expected to date material digitised through a different pipeline.

Licence

The source release states no licence, and none is set on this repository. The images came from Flickr under a mix of per-image rights; the project page retains copyright with the original holders and notes that a Creative Commons 2.0 licence applies only where individually identified. Check rights before redistributing or using commercially.

from datasets import load_dataset

ds = load_dataset("biglam/dating-historical-color-images", split="train")  # 1,326 rows

@inproceedings{palermo2012dating,
  author    = {Frank Palermo and James Hays and Alexei A. Efros},
  title     = {Dating Historical Color Images},
  booktitle = {Computer Vision -- ECCV 2012},
  series    = {Lecture Notes in Computer Science},
  volume    = {7577},
  pages     = {499--512},
  publisher = {Springer},
  year      = {2012},
  doi       = {10.1007/978-3-642-33783-3_36}}

Source data collected by Frank Palermo, James Hays and Alexei A. Efros; mirrored to the Hub by Daniel van Strien.