Here are some datasets that are relevant to my publications. The data can be used only for non-commercial purposes. Please cite the relevant publications if you use this data in your research.

IJCV 2012 Quantitative Evaluation

new

This dataset contains the 239 images which were used in the quantitative evaluation of our IJCV 2012 paper. The images were borrowed from the LabelMe dataset. The per-image annotations are available in XML format.

Webcam Clip Art dataset

The Webcam Clip Art Dataset contains images captured from high-resolution webcams all over the world, as introduced in our SIGGRAPH Asia 2009 paper.

ECCV 2010 Shadow Boundary Dataset

This dataset contains 135 images in which the ground shadow boundaries are labelled. These images were used in our ECCV 2010 paper for the training and testing of our shadow detection algorithm. Some of the images were taken from Flickr and LabelMe. Other images have been borrowed from Zhu et al., but annotations have been updated such that only ground shadow boundaries are labelled. In the case where the original images were high dynamic range (HDR), these images were cropped at 8 bits, and saved as jpegs.

  • Download (8.8MB)
  • Shadow annotations are saved in an XML format, for greater portability.
  • As convenience, MATLAB code for reading and displaying annotations is provided (look in the "src" folder).

ICCV 2009 Quantitative Evaluation

This dataset contains a subset of 391 images that were used to evaluate our single-image illumination estimation approach introduced in our ICCV 2009 paper. The XML files contain information about the sun position relative to the camera used as ground truth in our paper.

ICCV 2007 Composite images dataset

The semi-automatically generated dataset we used to obtain all the results in our ICCV 2007 paper is available for download. We also provide an example matlab function which loops over all the images in the database, extract various useful information, and displays it to the screen. This is meant to provide a starting point to anyone interested in performing experiments on the dataset.

Valid XHTML 1.0 Transitional