Indexed Dataset from YouTube for a Content-Based Video Search Engine

Abstract

Numerous researches on content-based video indexing and retrieval besides video search engines are tied to a large-scaled video dataset. Unfortunately, reduction in open-sourced datasets resulted in complications for novel approaches exploration. Although, video datasets that index video files located on public video streaming services have other purposes, such as annotation, learning, classification, and other computer vision areas, with little interest in indexing public video links for purpose of searching and retrieval. This paper introduces a novel large-scaled dataset based on YouTube video links to evaluate the proposed content-based video search engine, gathered 1088 videos, that represent more than 65 hours of video, 11,000 video shots, and 66,000 unmarked and marked keyframes, 80 different object names used for marking. Moreover, a state-of-the-art features vector, and combinational-based matching, beneficial to the accuracy, speed, and precision of the video retrieval process. Any video record in the dataset is represented by three features: temporal combination vector, object combination vector with shot annotations, and 6 keyframes, sideways with other metadata. Video classification for the dataset was also imposed to expand the efficiency of retrieval of video-based queries. A two-phased approach has been used based on object and event classification, storing video records in aggregations related to feature vectors extracted. While object aggregation stores video records with the maximal occurrence of extracted object/concept from all shots, event aggregation classify based on groups according to the number of shots per video. This study indexed 58 out of 80 different object/concept categories, each has 9 shot number groups.

YouTube Dataset on Mobile Streaming for Internet Traffic Modeling and Streaming Analysis

“…The datasets include, among others, DASH, H264 and H265, mobile video quality, or QoE datasets. For video streaming only, and in particular, several datasets already exist for the YouTube platform dealing with watch histories 27 , video application information, key-frame distribution and object names for search engines 28 . For viewing activity, in particular, Lall et al .…”

Development of an Effective Bootleg Videos Retrieval System as a Part of Content-Based Video Search Engine

“…Moreover, in an early experiment conducted in this study to find a threshold value that is optimal to our video index dataset [22], the tools involved in this experiment were PySceneDetect and FFmpeg. A deferential number of video files from the dataset were used and evaluated separately.…”

“…It is also crucial for video search with faster processing and less computational cost when it comes to commonly used frame-byframe matching. However, choosing keyframes is crucial with a maximal representation for every video shot and with no redundancy [22].…”

“…However, YOLOv3 which is based on TensorFlow framework was the main object detector for this work and experiments, as it was trained on Microsoft's COCO library using a diversity of objects types that reached 81 objects including added "null object" which represent frames that didn't return any objects by applying the object detector. A state-of-the-art framework called ImageAI [46] was used to implement object detection and keyframes annotations in both indexing and retrieval [22].…”

“…Moreover, this study offers a new YouTube crawled largescaled dataset created and based on the research work of [22], containing 1088 video records representing a sum of 65+ hours of video, extracting video shots of 113,502+ segments, and a total extracted keyframe images of 677,004+ marked/unmarked frames. In addition, object detection and marking the frames was done by using 80 different RetinaNet based objects trained using MS-COCO dataset [51] on top of ImageAI [46] platform, which is a state-of-the-art opensourced python built library that offers trained models that have been mined from ImageNet-1000 with 1000 diverse objects applied on platforms of ResNet, Inception-v3, DeneseNet, and SqueezeNet, and finally, imposing a video classification criterion on the indexed dataset for the purpose of increasing efficiency and decreasing the time of retrieving queries by example video.…”