Semantic video labeling by developmental visual agents

摘要

In the recent years, computer vision has been undergoing a period of great development, testified by the many successful applications that are currently available in a variety of industrial products. Yet, when we come to the most challenging and foundational problem of building autonomous agents capable of performing scene understanding in unrestricted videos, there is still a lot to be done. In this paper we focus on semantic labeling of video streams, in which a set of semantic classes must be predicted for each pixel of the video. We propose to attack the problem from bottom to top, by introducing Developmental Visual Agents (DVAs) as general purpose visual systems that can progressively acquire visual skills from video data and experience, by continuously interacting with the environment and following lifelong learning principles. DVAs gradually develop a hierarchy of architectural stages, from unsupervised feature extraction to the symbolic level, where supervisions are provided by external users, pixel-wise. Differently from classic machine learning algorithms applied to computer vision, which typically employ huge datasets of fully labeled images to perform recognition tasks, DVAs can exploit even a few supervisions per semantic category, by enforcing coherence constraints based on motion estimation. Experiments on different vision tasks, performed on a variety of heterogeneous visual worlds, confirm the great potential of the proposed approach.