Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.

#Photo-Art-50 dataset

The Photo-Art-50 dataset is a dataset of images from photos and artwork, with ground truth bounding boxes for 50 object classes. The aim is to evaluate cross-depiction object detection performance.

The dataset was originally produced by Qi Wu and Hongping Cai while working under Peter Hall at the University of Bath.

Please also see the People-Art dataset.


This dataset contains 50 classes of object. There are 90 to 138 images for each class, approximately half of which are photos and the other half art. The 50 classes all appear in Caltech-256. Some of the photos are from Caltech-256; the rest are from Google searches.

The artwork images came from searching using a variety of keywords to cover a wide gamut of depiction styles, e.g. "horse cartoon", "horse drawing", "horse painting", "horse sketches", "horse kid drawing", etc. All selected images have a reasonable size of a meaningful object area and there are ground-truth bounding boxes, labelled by hand, for each object.

Files included

  • {cls_id}.{cls_name}/{cls_id}.a_{img_id}.jpg: Art images for each class
  • {cls_id}.{cls_name}/{cls_id}.p_{img_id}.jpg: Photo images for each class
  • gt_bb/{cls_id}.txt: Ground truth file for each class

The ground truth files contain bounding boxes for each object instance as a row in ccv format: image_name x0 y0 w h NB image_name might appear more than once, which means there are multiple object instances in the image.


Many of the images are subject to copyright. These are provided only for "data mining for non-commercial research" or other "fair dealing" (UK Guidance).


The dataset appears in the following publications:

If you wish to cite the dataset, please use the following citation.


Photos and artwork images with object annotations for academic use only



No releases published


No packages published