Multi class object recognition with an adaptive confidence: Cascade of weak descriptors for fast hypothesis elimination

This paper points out the fact that object recognition methods are usually too complex for everyday life scenes. A robot helping humans in daily activities will need to recognize hundreds of different objects. In order to filter out unlikely models during recognition we propose the use of a cascade of simple visual descriptors. Our experiments use two global descriptors : spatial and color minimum volume bounding boxes. Results show this simple cascade can discard unlikely models up to 295 out of 300 instances and 50 out of 51 classes.