Interested Keyframe Extraction of Commodity Video Based on Adaptive Clustering Annotation

Abstract

Keyframe recognition in video is very important for extracting pivotal information from videos. Numerous studies have been successfully carried out on identifying frames with motion objectives as keyframes. The definition of “keyframe” can be quite different for different requirements. In the field of E-commerce, the keyframes of the products videos should be those interested by a customer and help the customer make correct and quick decisions, which is greatly different from the existing studies. Accordingly, here, we first define the key interested frame of commodity video from the viewpoint of user demand. As there are no annotations on the interested frames, we develop a fast and adaptive clustering strategy to cluster the preprocessed videos into several clusters according to the definition and make an annotation. These annotated samples are utilized to train a deep neural network to obtain the features of key interested frames and achieve the goal of recognition. The performance of the proposed algorithm in effectively recognizing the key interested frames is demonstrated by applying it to some commodity videos fetched from the E-commerce platform.