A Fully Dynamic Context Guided Reasoning and Reconsidering Network for Video Captioning.

Visual reasoning and reconsidering capabilities are instinctively executed alternately as people watch a video and attempt to describe its contents with natural language. Inspired by this, a novel network that joints fully dynamic context guided reasoning and reconsidering is proposed in this paper. Specifically, an elaborate reconsidering module referred to as the reconsiderator is employed for rethinking and sharpening the preliminary results of stepwise reasoning from coarse to fine, thereby generating a higher quality description. And in turn, the reasoning capability of the network can be further boosted under the guidance of the context information summarized during reconsidering. Extensive experiments on two public benchmarks demonstrate that our approach is pretty competitive with the state-of-the-art methods.