Computer Vision News - January 2018

Method: In RCN, objects are modeled from combination of contours and surfaces . Surfaces are modeled using a conditional random field (CRF) . Contours are modeled using a compositional hierarchy of features . This factored and flexible contour and surface representation enables the model to classify and detect various objects without training on every surfaces and contours combination. RCN is a hierarchical model and this hierarchy fulfils two roles: 1) enabling the representation of deformations through multiple levels; 2) efficient sharing of features between different objects. While most CNN models developed in recent years, first, analyze whole images, and second, presume very little knowledge about the images and objects in them, RCN aims at modeling the contours and surfaces of objects and background in the image in a direct way. The figure below illustrates the structure and internal workings of RCN. On the left is a toy example of a three level RCN representation of a square, with level 1 representing lines, and level 2 representing the four corners -- you can see the lateral connections within the network, as well as the pooling layers between levels of the hierarchy. On the right is a realistic four level RCN representation of the letter “ A ”, built with the same structure. The implementation consists of two steps: 1) PreProc -- a set of filters that extract low-level elements from the image. 2) Learns a hierarchical model from the data. 6 Computer Vision News Research Research “RCN’s performance suggests that incorporating insights from neuroscience can lead to highly data-efficient, generalizable and robust machine-learning models.”