Back to Guides

    15 Video Datasets for Computer Vision, Compared (2026)

    One table: size, labels, footage and licence for the video datasets computer-vision teams actually use, and what none of them covers.

    By Simuletic
    July 4, 2026 · Updated October 24, 2026
    7 min read

    In short: for pre-training a video backbone, use Kinetics or Something-Something. For activities in context, AVA and Charades. For skeletons, NTU RGB+D. For surveillance, MEVA is the only large CCTV-style set with a commercial-friendly licence (CC BY 4.0); UCF-Crime, XD-Violence and RWF-2000 have the incidents but restrictive or unclear terms. None has many examples of falls, weapons or theft from your own camera angle.

    This page is the comparison. For how to label activity data and build a HAR dataset, read the human activity recognition guide; for incident detection on CCTV, the video anomaly detection guide.

    High CCTV view of a stadium concourse in the rain with three people fighting, a fight box around the group and pose skeletons on each person
    A fight seen from a high CCTV camera, the kind of footage most public video sets lack.

    The comparison table

    Sizes are as published; YouTube-based sets lose clips over time as videos are removed. "Check the project page" means we could not find a licence statement to verify: treat it as not cleared for commercial use until you have.

    DatasetSizeLabelsFootageLicence
    Action recognition
    UCF101 (2012)13,320 clips101 action classes, one per clipYouTube, mostly handheldNone stated; clips are third-party YouTube videos
    HMDB51 (2011)~6,800 clips (papers report 6,766–6,849)51 action classesFilms and web videoCheck the project page
    Kinetics-400 (2017)~306K clips of ~10 s at release; fewer still online400 action classesYouTube, mostly handheldNone in the official readme; third-party YouTube content
    Kinetics-700 (2019)~650K clips of ~10 s700 action classes, ≥600 clips eachYouTube, mostly handheldNone in the official readme; third-party YouTube content
    Something-Something v2220,847 clips174 hand–object interaction classesCrowd-workers filming themselvesQualcomm licence; check its terms
    Activities in context
    AVA Actions (2018)430 movie clips of 15 min80 atomic actions, boxed per person (spatio-temporal)Feature filmsAnnotations CC BY 4.0 (Google); the films are copyrighted
    Charades (2016)9,848 videos157 action classes, 66,500 temporal annotationsCrowd-workers acting at homeCheck the licence file in the download
    Skeleton / multi-modal
    NTU RGB+D 60 / 120 (2016 / 2019)56,880 / 114,480 samples60 / 120 classes incl. falling and mutual actionsLab, three Kinect v2 sensors: RGB, depth, IR, 3D skeletonsAcademic research only; commercial use needs permission
    Surveillance & anomaly
    MEVA (2020)9,300+ h recorded; ~22 h of annotations public37 activity types38 RGB and thermal IR cameras at one venue, plus UAVCC BY 4.0
    UCF-Crime (2018)1,900 untrimmed videos, ~128 h13 anomaly classes + normal; video-level training labelsReal CCTV from YouTube and LiveLeakResearch use only
    XD-Violence (2020)4,754 videos, 217 h, with audio6 violence classes + normalMovies and YouTube: CCTV, handheld, dashcamCheck the project page
    RWF-2000 (2019)2,000 clips of up to 5 sViolent / non-violentSurveillance footage from YouTubeCommercial use and redistribution need the lab's approval
    ShanghaiTech Campus437 videos, 130 abnormal eventsOne-class (normal vs. abnormal), pixel masksFixed campus cameras, 13 scenesCheck the project page
    CUHK Avenue (2013)37 clips, 47 abnormal eventsOne-class, boxed abnormal eventsOne fixed camera on a walkwayCheck the project page
    UBnormal (2022)543 videos, 29 scenesOpen-set anomalies, pixel masksSynthetic people on real background photosCC BY-NC-ND 4.0

    Papers: Soomro et al. UCF101; Kay et al. Kinetics; Carreira et al. Kinetics-700; Gu et al. AVA; Sigurdsson et al. Charades; Liu et al. NTU RGB+D 120; Corona et al. MEVA; Sultani et al. UCF-Crime; Wu et al. XD-Violence; Cheng et al. RWF-2000; Acsintoae et al. UBnormal.

    How to pick

    • Pre-training a video model: Kinetics-400/700 for appearance, Something-Something v2 when motion direction matters (pushing vs. pulling).
    • Who is doing what, where: AVA (boxes per person per second) and Charades (several overlapping activities per video).
    • Skeleton-based recognition: NTU RGB+D 60/120, the standard for graph-convolution models; research-only.
    • Surveillance: MEVA for licence-clean CCTV-style activity; UCF-Crime and XD-Violence to compare against published anomaly results; ShanghaiTech and Avenue for one-class methods.
    • Shipping a product: read the licence column first. Most of this table is research-only or unclear.
    Overhead CCTV view of a phone shop where a woman hides a boxed phone under her shirt, with person boxes and pose keypoints
    A frame from the shoplifting dataset, which also ships video sequences.

    What none of them covers

    The public sets were built to compare models, not to ship safety products. Three gaps show up in every CCTV project we see:

    1. Rare events in volume. UCF-Crime, one of the largest incident sets, has 50 shoplifting and 50 fighting videos. A deployed model needs many more examples, from varied scenes.
    2. Your camera angle. Most clips are handheld, film or lab footage. A camera mounted high on a wall or ceiling changes how a fall, a knife or a concealed item looks.
    3. Look-alikes. None of the sets deliberately labels the things that trigger false alarms: someone sitting on the floor, an umbrella, a hug.
    Strip of CCTV frames of people holding crisps, a hair dryer and umbrellas, each labelled no weapon
    Labelled look-alikes from the weapon dataset: everyday objects that should not raise an alarm.

    That is the gap our synthetic CCTV datasets fill. Each is seen from real CCTV positions, labelled, and includes look-alike negatives:

    The catalogue datasets are labelled images (1,000 per licence); the shoplifting dataset includes video. Video clips with per-frame boxes, tracking IDs and event start/end times are made to order on your own camera views; see custom data.

    FAQ

    What is the largest public video dataset for action recognition?

    Of the sets in this table, Kinetics-700 is the largest labelled action set at about 650,000 ten-second YouTube clips in 700 classes. MEVA has more raw footage (over 9,300 hours) but only about 22 hours of public annotations.

    Which video datasets can I use commercially?

    Few. MEVA is CC BY 4.0 and AVA's annotations are CC BY 4.0 (the films are not). NTU RGB+D and UCF-Crime are research-only, UBnormal is non-commercial, RWF-2000 needs the lab's approval, and the YouTube-based sets leave the video rights with the uploaders.

    Which public video dataset is closest to real CCTV?

    MEVA (fixed RGB and thermal cameras at one venue, CC BY 4.0), UCF-Crime (real CCTV incidents, research-only) and RWF-2000 (surveillance fights) are closest. ShanghaiTech and CUHK Avenue are fixed cameras too, but on a single campus.

    Do I need video, or are images enough?

    Most CCTV alerting runs a detector per frame plus simple rules over time, so labelled images from the right camera angle carry most of the value. Temporal models (3D CNNs, video transformers, skeleton sequences) need clips with start and end times.

    Does Simuletic sell video datasets?

    The catalogue is mostly labelled image datasets; the shoplifting dataset includes video sequences and a few sets add evaluation clips. Video with per-frame boxes, tracking IDs and event start/end times is made to order on your own camera views.

    Need a video dataset that doesn't exist yet?

    Send us the behaviour, a few stills of your empty camera views and the label format. We make synthetic clips with per-frame labels, quoted per project.