Adaptable Neural Networks for Unsupervised Video Object Segmentation of Stereoscopic Sequences

Abstract

An adaptive neural network architecture is proposed in this paper, for efficient video object segmentation and tracking in stereoscopic sequences. The scheme includes: (A) A retraining algorithm that optimally adapts the network weights to the current conditions and simultaneously minimally degrades the previous network knowledge. (B) A semantically meaningful object extraction module for constructing the retraining set of the current conditions and (C) a decision mechanism, which detects the time instances when network retraining is required. The retraining algorithm results in the minimization of a convex function subject to linear constraints. Furthermore description of the current conditions is achieved by appropriate combination of color and depth information. Experimental results on real life video sequences indicate the promising performance of the proposed adaptive neural network-based video object segmentation scheme.