Integrating Structural Information in Deep Convolutional Neural Networks for Low- and High-Level Vision

Iasonas Kokkinos
(Ecole Centrale Paris, France)

Abstract:

Deep Convolutional Neural Networks (DCNNs) have demonstrated excellent performance in a large variety of low- and high-level tasks, but they are largely considered as `black box' classifiers that do not exploit the insights we have about the computer vision task. In this talk we will present recent efforts on incorporating classical computer vision ideas in DCNNs used to addressed both low- and high- level tasks.

We will start with scale-invariance in image labelling and image classification, move on to treating scale- and aspect- variation in object detection, and finally turn to fully-connected models for semantic image segmentation and boundary detection. The common theme in all these works is the use of fully convolutional neural networks, aimed at substituting the commonly hand-engineered pipelies as generic visual front-ends.

Our semantic segmentation system is publicly available from this link: https://bitbucket.org/deeplab/deeplab-public/