Computer Vision News - March 2017

Computer Vision News Tool 11 Faster R-CNN: The third network of the series to be developed was called Faster R-CNN. In this model there is a single, unified network for object detection, formed by including the region proposal element in the network itself as part of the training process, rather than as an external element. This element was called the Region Proposal Network or RPN – this module serves as the ‘attention’ of the network. The figure below shows the structure of the RPN module in more detail: At each sliding-window location, the module produces multiple region proposals; number of proposals per location is denoted as k . The region element (denoted as reg layer ) has 4 outputs which are the coordinates of the proposal box, and the category classifier (denoted as cls layer ) predicts the probability of a region proposal containing an object or not. In the following image, the RPN in its proper position within the Faster R-CNN network: To sum up, in Faster R-CNN we have 4 elements to train simultaneously. The first two elements are a learning external region proposal network that replaces the slow selective search; the last two elements are the ‘original’ region and category classifier elements of the R-CNN. Tool