This is a test dataset for trainable background subtraction algorithms. It shows a complex scene under various lighting conditions and several objects in the foreground that are manually presented to the camera. (about 500 training images and 90 test images with ground truth segmentations at a resolution of 320×240 pixels).

You can download the dataset here (file size: 100MB).

If you have questions, please contact Adrian Ulges.

The dataset contains 100 camera captured documents with their corresponding ASCII text ground-truth. 10 sample documents from the dataset are available here.

warped-documents-samples.zip

In addition to the ASCII text ground-truth, text-line and zone level ground-truth is also provided. The ground-truth is represented in the color-channels of the documents as described in:

Faisal Shafait, Daniel Keysers, Thomas M. Breuel, ”Pixel-Accurate Representation and Evaluation of Page Segmentation in Document Images”, ICPR 2006, International Conference on Pattern Recognition, pages 872-875

Original grayscale documents can be obtained upon request. This dataset was used for the document image dewarping competition held with CBDAR 2007. Complete dataset with the corresponding ground-truth can be downloaded from the contest webpage. If you have any questions, please contact Faisal Shafait.

The dataset contains 25 scanned images of Urdu text. The images are categorized into five classes: book, novel, poetry, magazine, and newspaper. There are 5 images of each class in the dataset.

urdu-documents-book.zip

urdu-documents-magazines.zip

urdu-documents-novel.zip

urdu-documents-poetry.zip

urdu-documents-newspaper.zip

This dataset provides 180 videos showing 15 objects. It serves as a testbed for the training and recognition of object appearance in video streams, with the focus on a real-world object recognition scenario including clutter, motion blur, and illumination changes.

The videos were generated by choosing 15 objects and manually presenting them to a web-cam (UniBrain Fire-I) at a framerate of 25/s. For each object, 12 videos were taken, each providing about 40 frames at a 320×240 resolution.

While the objects and the arm of the operator move, the background remains static. The videos were taken at two different locations providing a different background and illumination:

1) office: Frames are generally darker. and the background is simple except for a few strong edges.

2) lab: The scenes contain strong daylight and many specular highlights. The background is strongly textured.

Though a front side is chosen for each object, slight pose changes occur due to the manual presentation. Also, a strong motion blur can be observed in many frames.

For more information and download, goto the VideoObjects Website.