Abstract
This paper introduces the problem of Fine-grained Incident Video Retrieval (FIVR). Given a query video, the objective is to retrieve all associated videos, considering several types of association that range from duplicate videos to videos from the same incident. FIVR offers a single framework that contains as special cases several retrieval tasks. To address the benchmarking needs of all such tasks, we constructed and present a large-scale annotated video dataset, which we call FIVR-200K and it comprises 225,960 videos. To create the dataset, we devised a process for the collection of YouTube videos based on major events from recent years crawled from Wikipedia and deployed a retrieval pipeline for the automatic selection of query videos based on their estimated suitability as benchmarks. We also devised a protocol for the annotation of the dataset with respect to the four types of video association defined by FIVR. Finally, we report results of an experimental study on the dataset comparing a variety of state-of-the-art visual descriptors and aggregation techniques, highlighting the challenges of the problem at hand.
Abstract (translated)
本文介绍了细粒度事件视频检索（FIVR）的问题。给定查询视频，目标是检索所有相关视频，考虑从重复视频到同一事件的视频的几种类型的关联。 FIVR提供了一个单独的框架，其中包含几个检索任务的特殊情况。为了满足所有此类任务的基准测试需求，我们构建并呈现了一个大型带注释的视频数据集，我们称之为FIVR-200K，它包含225,960个视频。为了创建数据集，我们设计了一个基于近几年来自维基百科的主要事件收集YouTube视频的流程，并根据其估计的适合性作为基准，部署了一个检索管道，用于自动选择查询视频。我们还设计了一个关于FIVR定义的四种视频关联的数据集注释协议。最后，我们报告了数据集的实验研究结果，比较了各种最先进的视觉描述符和聚合技术，突出了手头问题的挑战。
URL
https://arxiv.org/abs/1809.04094
