Region detection and description for Object Category Recognition

The way images are decomposed and represented biases how well subsequent object learning and recognition methods will perform. We choose to initially represent the images by sets of local distinctive regions and their description vectors. We evaluate the problems of distinctive region detection and description in two separate stages, by first reviewing some of the state-of-the-art methods, and then discussing the methods we propose to use for object category recognition. In comparing the performance of our region detection-description technique for scale and rotation invariance with the performance of the other detection-description techniques, we find that our approach provides better results than existing methods, in the context of object category recognition. The evaluation consists of clustering similar descriptor regions and computing (1) the number of single measure clusters (measures intra-class sensitivity), (2) cluster precision clusters (measures how clusters are shared between different classes) and (3) the generalizability property of regions (measures matching to classes). Our technique, which is a variant on the Kadir-Brady saliency detector scored better and not worse than all the other methods evaluated.