'Affordance' detection by mid-level physical parts

Derived from ecological psychology, the term 'affordance' refers to the functional classification of objects. It simply means the set of actions a subject (i.e. humans and anthropomorphic agents) can possibly perform with an object. There are several paradigms in the researches regarding the approaches of affordance detection. These approaches include considering other contexts like the subject, ambient environment, scene etc. along with the object to detect the affordance. But the core challenge, which is to identify the affordance of an object with mere its visual features is still open ended. The orthodox approaches often used the fine level image features (i.e. HOG, SIFT) and perform bottom up processing to detect affordance. In this paper, we propose an effective methodology of using 'mid-level physical parts' of objects as features for affordance detection. It is a multi-layered detection problem, where the simple image features are used to detect mid-level physical parts of an object and subsequently those parts are used to detect affordance. Our experiment results show the effectiveness of our method over the current affordance detection techniques. Moreover since the parts may be shared between different object classes, the method is also proficient in generality.