The Evolution of Object Categorization and the Challenge of Image Abstraction

Using surfaces and surface relations in an Early Cognitive Vision system

“…As Dickinson and others (see, e.g., [13,46]) pointed out, the focus of the computer vision community has-after realizing severe problems in approaching computer vision in a hierarchical paradigm (as in particular suggested by Marr [53])-shifted from explicitly designed hierarchies to the design of more efficient low level feature descriptors and classification schemes based on those leading to flat architectures. We argued in [46] that at the time Marr published his approach towards computer vision, two main reasons made his ideas unfeasible.…”

“…In [13], Dickinson gives an overview of the development of computer vision approaches pointing to the 'representational gap' between sensory information and the categorical models applied. Dickinson argues that while in the 1970's there existed a large gap between the degree of abstraction used in the applied models and the features extracted from the input image, this gap has been reduced in the 1980's and 1990's by reducing the complexity of the categorical models and to a certain extend also by progress of feature extraction algorithms.…”

“…Interactions between the disciplines of "biological vision" and "computer vision" have varied in intensity throughout the course of computer vision history and have in some way reflected the changing research focuses of the machine vision community (see, e.g., [13]). Without any doubt, the groundbreaking work of Hubel and Wiesel [28] gave a significant impulse to the computer vision community via Marr's work [53] on building visual hierarchies analogous to the primate visual system.…”

Object Categorization in Clutter Using Additive Features and Hashing of Part-Graph Descriptors

“…In order to match the perception capabilities of humans, Dickinson in [2] advocates that searching for predefined templates is not enough, and that recognition of new exemplars of known categories has to be facilitated. On this premise, Marton et al [3] use geometric cues for categorization and visual cues for instance classification.…”

“…Nonetheless, no matter how extensive the training data, a robot might always be confronted with a novel object instance or type when operating in a real dynamic environment. Since teaching all the possible objects a robot might encounter is not always possible, categorization becomes an important step towards learning these novel objects [2]. Thus a mechanism is needed to enable a robot to autonomously acquire new object models as it operates in an environment, and to efficiently add it to its classification framework [3].…”