Analyzing Appearance and Contour Based Methods for Object Categorization
This set contains the contours for all objects as binary images (in the original resolution). We used it in our local shape experiments.
All images are cropped, so that they contain only the object without any border area. In addition, they are rescaled to a size of 128*128 pixels. Again, the scale is left the same for all images of the same object. This dataset is useful when no derivatives need to be calculated. We used it in the PCA experiments.
The images of the ETH-80 database in their original resolution (ranging from 400*400 to 700*700 pixels, depending on object size). All images are cropped, so that they contain only the object, centered in the image, plus a 20% border area. Due to space constraints, this dataset is not included on the general webpage, but will be made available upon request.
This set also contains all images, rescaled to size 128*128, but this time in the classic COIL style. That is, the scale is chosen per image, such that the object's bounding box always fills the complete image. This may lead to rather large scale changes for different views of the same object, but we chose to include it for compatibility reasons.
This is the standard set of the database, used for almost all experiments. All images are cropped, so that they contain only the object, centered in the image, plus a 20% border area (to avoid border effects when derivatives need to be calculated). In contrast to the COIL database, the scale is left the same for all images of the same object. Images are rescaled to a size of 256*256 pixels.