Deep neural networks (DNNs) are typically trained on specific datasets, optimized with particular discriminating capabilities. Often several different DNN topologies are developed solving closely related aspects of a computer vision problem. But to utilize these topologies together, leveraging their individual discriminating capabilities, requires implementing each DNN separately, increasing the cost of practical solutions.
In this talk, a methodology to merge multiple deep networks using graph contraction is developed. This provides a single network topology, achieving significant reduction in size over the individual networks. More significantly, this merged SPDNN network can be re-trained across the combined datasets used to train the original networks, improving its accuracy over the original networks. The result is a single network that is more generic, but with equivalent – or often enhanced – performance over a wider range of input data.
Examples of several problems in contemporary computer vision are solved using SPDNNs. These include significantly improving segmentation accuracy of eye-iris regions (a key component of iris biometric authentication) and mapping depth from monocular images, demonstrating equivalent performance to stereo depth mapping.