Developmental Learning for Object Perception

The computer vision community provides a large amount of object recognition approaches which are mostly based on prior knowledge, either in the form of algorithm choices or image databases. In contrast, our algorithm is designed by analogy with the human perceptual learning in terms of input data and learning organisation. It does not require image databases nor face/skin detectors, and allows to be robust to changes in the environment, background, lighting conditions and camera motion. We acquire all knowledge iteratively from low-level features. Our approach differs from Deep Belief Networks and Deep Energy Models [2] in that it does not use training and testing sets, but characterizes each object by a hierarchy of shared and complementary features (see Fig. 1) built up incrementally, during interactions with objects.