Gestalt-inspired features extraction for object category recognition

We propose a methodology inspired by Gestalt laws to extract and combine features and we test it on the object category recognition problem. Gestalt is a psycho-visual theory of Perceptual Organization that aims to explain how visual information is organized by our brain. We interpreted its laws of homogeneity and continuation in link with shape and color to devise new features beyond the classical proximity and similarity laws. The shape of the object is analyzed based on its skeleton (good continuation) and as a measure of homogeneity, we propose self-similarity enclosed within shape computed at super-pixel level. Furthermore, we propose a framework to combine these features in different ways and we test it on Caltech 101 database. The results are good and show that such an approach improves objectively the efficiency in the task of object category recognition.