Shifting More Attention to Video Salient Object Detection
A complete, high-quality video marked a significant target detection (video salient object detection, VSOD) data sets in real life is missing. Thus, a video attention constructed herein consistent, densely labeled VSOD data set, referred DAVSOD. This dataset contains 226 videos were 23,938, covering a variety of real-world scenarios, the object, and examples of operation. By means of corresponding real eye point of gaze data, user accurate labeling. Thus for the first time explicitly emphasized the significant transfer phenomenon challenging that video significantly objects may be changed dynamically.
In order to provide a comprehensive evaluation, the system described herein to assess the art VSOD 17 VSOD most representative algorithm in seven existing data set and data set DAVSOD constructed. At the same time, provides a benchmark model, it is equipped with a memory length of time the network convolution of a significant shift for the (convLSTM), can be effective dynamic video capture significant attention by studying human behavior.
This article contributions:
1) construct a data set;
2) 17 of the most advanced model of comprehensive evaluation, at the same time, propose a basic model called the SSAV.
Base model proposed frame structure is as follows:
