Final Report Summary - VIDEOLEARN (Video and 3D Analysis for Visual Learning)
A major challenge in computer vision these days is how to learn effective visual representations of the environment that can be used for decision making and planning. In this project, we focused particularly on learning representations from videos. We investigated how to obtain large amounts of annotated data and developed efficient interactive video segmentation solutions for manual annotation. Moreover, we found an effective way of data acquisition via automated rendering of images and annotation from synthetic scene models. We investigated how representations for motion and 3D structure can be learned end-to-end from training data and found very effective solutions that generalize to images that are very different from the training data. The project established deep learning methods for motion estimation, for 3D object modeling, and for two-frame structure from motion.