Computer vision applications commonly using data sets for training object recognition systems, sets that include images which are tightly cropped around objects. This approach decreases the running time and increase recognition accuracy by providing best characteristics of the objects of interests to the learning algorithms. In an effort to create best test set for learning algorithms, we decided to use superpixel grouping based object recognition. Resulting objects can be used for training the learning algorithms. While this approach can be efficient for complex recognition algorithms which will focus only proposed objects from our algorithm rather than calculating the objects by itself, it can also be powerful and efficient for other but rather simple computer vision applications as well.
You can find resulting paper under the Paper folder.