Source-linked AI summary
V3C - a Research Video Collection
Luca Rossetto, Heiko Schuldt, George Awad, Asad A. Butt
TL;DR
Large web-video collections have grown, but existing research datasets may not reflect the content commonly found online. The paper introduces V3C, a collection of Vimeo Creative Commons videos with shot and metadata resources, for video retrieval research and related applications.
Problem
Existing video datasets may be too small or may not reflect Internet video in content, length, or resolution, despite the need for standardized research datasets.
Source: §1 · p. 1Method
Results
V3C comprises 28,450 Vimeo videos with shot boundaries, keyframes, and metadata, totaling roughly 3,800 hours.
Source: §1 · p. 2, §4 · p. 11Takeaways & Limitations
The collection’s content diversity and varied resolutions make it useful for video-related applications, including transport and storage research and shot-level video captioning.
Source: §3.3 · pp. 10–11Collection sourcing is bounded to freely accessible platforms with diverse contemporary video and some Creative Commons-licensed material.
Source: §2 · p. 2
Abstract
from arXiv · showhide
With the widespread use of smartphones as recording devices and the massive growth in bandwidth, the number and volume of video collections has increased significantly in the last years. This poses novel challenges to the management of these large-scale video data and especially to the analysis of and retrieval from such video collections. At the same time, existing video datasets used for research and experimentation are either not large enough to represent current collections or do not reflect the properties of video commonly found on the Internet in terms of content, length, or resolution. In this paper, we introduce the Vimeo Creative Commons Collection, in short V3C, a collection of 28'450 videos (with overall length of about 3'800 hours) published under creative commons license on Vimeo. V3C comes with a shot segmentation for each video, together with the resulting keyframes in original as well as reduced resolution and additional metadata. It is intended to be used from 2019 at the International large-scale TREC Video Retrieval Evaluation campaign (TRECVid).
1 Introduction
Despite the growth of web video, standardized research datasets remain needed, and existing collections differ from video commonly found online. The paper presents V3C as a collection intended to better represent Internet video in retrieval evaluation campaigns.
1 Introduction
- Existing video collections differ substantially from the type of web video commonly found “in the wild.”The paper identifies the need for standardized datasets despite the availability of large amounts of video.Source: §1 · pp. 1–2
- V3C contains 28,450 Vimeo videos with metadata, shot segmentation, and representative keyframes at original and reduced resolutions.The collection is intended to complement or replace existing collections in real-world video retrieval evaluation campaigns.Source: §1 · p. 2
2 Collection Process
The collection process selected a platform and video categories to meet real-world and redistribution requirements, then filtered, checked, and partitioned the videos.
2 Collection Process
- The source had to be freely accessible, host diverse contemporary video, and include content published under a Creative Commons license for redistribution.Source: §2 · p. 2
- Vimeo was chosen over YouTube because YouTube’s Terms of Service forbid downloading videos except for playback in a video stream.Source: §2 · p. 2
- The selection included four complete top-level categories and selected subcategories, excluding content with low visual diversity, animation, or little describable activity.Personal, Documentary, Sports, and Travel were included in full; selected videos were then filtered by duration and license.Source: §2 · pp. 2–3
- Candidate videos were downloaded sequentially and checked for successful decoding.Source: §2 · p. 3
- Cineast segmentation flagged videos with unusually different segment-length distributions for manual inspection, and the collection was checked for exact duplicates.Source: §2 · p. 3
- Three randomly selected subsets of increasing size received sequential numeric IDs so multiple parts could be used together.Source: §2 · p. 3
3 The Vimeo Creative Commons Collection
V3C contains 28,450 videos organized into three partitions, with shot boundaries, keyframes, thumbnails, and metadata. Its reported properties broadly resemble Vimeo’s wider video collection, and its varied content and resolutions support several research uses.
3.1 Collection Structure
- V3C contains 28,450 videos lasting 3–60 minutes each, divided into three partitions with shot boundaries, representative keyframes, thumbnails, and video metadata.Keyframes retain original resolution; thumbnails are downscaled to 200 pixels wide.Source: §3.1 · p. 3
3.2 Statistical Properties
- The upload-age distributions resemble each other where data overlap, while V3C includes videos uploaded as late as early 2018.The comparison dataset in [7] was completed in mid-2016.Source: §3.2 · p. 5, Figure 3 · p. 7
- Where unrestricted, V3C’s video-duration and resolution distributions follow the broader Vimeo distributions rather closely.The paper therefore describes V3C as reasonably representative of Vimeo web video on these properties.Source: §3.2 · p. 5, Figure 4 · p. 7, Figure 5 · p. 8
- The detected title-and-description language distribution is similar to the Vimeo dataset for videos where language detection succeeds; audio was not analyzed.Table 2 reports the top ten languages for either dataset and marks failed detection as “?”.Source: §3.2 · p. 5, Table 2 · p. 5
3.3 Possible Uses
- V3C’s varied content and resolutions make it relevant to machine learning, computer vision, and video transport and storage applications.The paper also identifies video tagging, ad-hoc search, known-item search, and captioning as possible retrieval applications.Source: §3.3 · pp. 10–11
3.4 Availability
- The collection was planned for release to 2019 TRECVid and VBS participants and the public, with judgments and queries/topics to follow each benchmark cycle.The post-cycle materials were intended to support reuse in local experiments and reproduction of results.Source: §3.4 · p. 11
4 Conclusions
V3C is a roughly 3,800-hour collection of Creative Commons video from Vimeo, augmented with metadata, shot boundaries, and keyframes. Its three partitions increase in length for use in video-search benchmarks over at least three consecutive years.
- V3C contains roughly 3,800 hours of Vimeo Creative Commons video, with technical and semantic metadata, shot boundaries, keyframes, and three partitions increasing from roughly 1,000 to 1,500 hours for benchmarks over at least three years.Source: §4 · p. 11