Query-Driven Video Summarization for Long Video Footage Analysis Using Faster-RCNN and Determinantal Point Processes

With ever growing volume of video content, video summarization becomes crucial for efficiently condensing lengthy videos into informative representations. This paper introduces a novel approach for video summarization by combining object identification with Determinantal Point Processes (DPP). Object identification, powered by advanced computer vision techniques, tracks and recognizes objects across video frames (Convolutional Neural Network/Computer Vision), allowing for identification and prioritization of frames with key objects and respective interactions. The proposed algorithm leverages DPP and R-CNN concurrently to select frames while avoiding redundancy and preserving diversity Extensive experiments using diverse video datasets demonstrates that this method performs competitively with existing summarization techniques while tested on TVSum dataset [12], a popular baseline. The results highlight the efficacy of the proposed algorithm in striking a balance between information retention and summary length. It provides accuracy of 84.34% and recall of 13.09%. This research contributes to the field of multimedia content analysis, with potential applications in video indexing, retrieval, and content recommendation systems. Through integration of object identification and DPP, this work introduces an innovative dimension to video summarization. The proposed approach holds promise for generating more informative and visually coherent video summaries. This work holds the potential to advance the state-of-the-art in video summarization techniques, benefiting various multimedia applications by aiding content analysis and user experience.