Video

MPI (Max Planck Institute) Sintel is a dataset for optical flow evaluation that has 1064 synthesized stereo images and ground truth data for disparity. Sintel is derived from open-source 3D animated short film Sintel. The dataset has 23 different scenes. The stereo images are RGB while the disparity is grayscale. Both have resolution of 1024×436 pixels and 8-bit per channel. Source: Fast Disparity Estimation using Dense Networks*

YouCook2 is the largest task-oriented, instructional video dataset in the vision community. It contains 2000 long untrimmed videos from 89 cooking recipes; on average, each distinct recipe has 22 videos. The procedure steps for each video are annotated with temporal boundaries and described by imperative English sentences (see the example below). The videos were downloaded from YouTube and are all in the third-person viewpoint. All the videos are unconstrained and can be performed by individual persons at their houses with unfixed cameras. YouCook2 contains rich recipe types and various cooking styles from all over the world. Source: http://youcook2.eecs.umich.edu/ Image Source: https://competitions.codalab.org/competitions/20594

The Moving MNIST dataset contains 10,000 video sequences, each consisting of 20 frames. In each video sequence, two digits move independently around the frame, which has a spatial resolution of 64×64 pixels. The digits frequently intersect with each other and bounce off the edges of the frame Source: Mutual Suppression Network for Video Prediction using Disentangled Features Image Source: http://www.cs.toronto.edu/~nitish/unsupervised_video/

The MOTChallenge datasets are designed for the task of multiple object tracking. There are several variants of the dataset released each year, such as MOT15, MOT17, MOT20. Source: MOTChallenge 2015: Towards a Benchmark for Multi-Target Tracking

The Breakfast Actions Dataset comprises of 10 actions related to breakfast preparation, performed by 52 different individuals in 18 different kitchens. The dataset is one of the largest fully annotated datasets available. The actions are recorded “in the wild” as opposed to a single controlled lab environment. It consists of over 77 hours of video recordings. Source: https://serre-lab.clps.brown.edu/resource/breakfast-actions-dataset/ Image Source: https://serre-lab.clps.brown.edu/resource/breakfast-actions-dataset/

The UCY dataset consist of real pedestrian trajectories with rich multi-human interaction scenarios captured at 2.5 Hz (Δt=0.4s). It is composed of three sequences (Zara01, Zara02, and UCY), taken in public spaces from top-view. Source: Trajectron++: Dynamically-Feasible Trajectory Forecasting With Heterogeneous Data Image Source: http://trajnet.stanford.edu/data.php?n=1

LAION-400M is a dataset with CLIP-filtered 400 million image-text pairs, their CLIP embeddings and kNN indices that allow efficient similarity search. #### ⚠️ Disclaimer & Content Warning (from the authors) Our filtering protocol only removed NSFW images detected as illegal, but the dataset still has NSFW content accordingly marked in the metadata. When freely navigating through the dataset, keep in mind that it is a large-scale, non-curated set crawled from the internet for research purposes, such that collected links may lead to discomforting and disturbing content. Therefore, please use the demo links with caution. You can extract a “safe” subset by filtering out samples drawn with NSFW or via stricter CLIP filtering. There is a certain degree of duplication because we used URL+text as deduplication criteria. The same image with the same caption may sit at different URLs, causing duplicates. The same image with other captions is not, however, considered duplicated. Using KNN clustering should make it easy to further deduplicate by image content.

This paper introduces the pipeline to scale the largest dataset in egocentric vision EPIC-KITCHENS. The effort culminates in EPIC-KITCHENS-100, a collection of 100 hours, 20M frames, 90K actions in 700 variable-length videos, capturing long-term unscripted activities in 45 environments, using head-mounted cameras. Compared to its previous version (EPIC-KITCHENS-55), EPIC-KITCHENS-100 has been annotated using a novel pipeline that allows denser (54% more actions per minute) and more complete annotations of fine-grained actions (+128% more action segments). This collection also enables evaluating the "test of time" - i.e. whether models trained on data collected in 2018 can generalise to new footage collected under the same hypotheses albeit "two years on". The dataset is aligned with 6 challenges: action recognition (full and weak supervision), action detection, action anticipation, cross-modal retrieval (from captions), as well as unsupervised domain adaptation for action recognition. For each challenge, we define the task, provide baselines and evaluation metrics.

The Kinetics-600 is a large-scale action recognition dataset which consists of around 480K videos from 600 action categories. The 480K videos are divided into 390K, 30K, 60K for training, validation and test sets, respectively. Each video in the dataset is a 10-second clip of action moment annotated from raw YouTube video. It is an extensions of the Kinetics-400 dataset. Source: Learning to Localize Actions from Moments Image Source: https://towardsdatascience.com/downloading-the-kinetics-dataset-for-human-action-recognition-in-deep-learning-500c3d50f776

The YouTube-8M dataset is a large scale video dataset, which includes more than 7 million videos with 4716 classes labeled by the annotation system. The dataset consists of three parts: training set, validate set, and test set. In the training set, each class contains at least 100 training videos. Features of these videos are extracted by the state-of-the-art popular pre-trained models and released for public use. Each video contains audio and visual modality. Based on the visual information, videos are divided into 24 topics, such as sports, game, arts & entertainment, etc Source: Audio-Visual Embedding for Cross-Modal Music Video Retrieval through Supervised Deep CCA

The SumMe dataset is a video summarization dataset consisting of 25 videos, each annotated with at least 15 human summaries (390 in total). Source: https://gyglim.github.io/me/vsum/index.html Image Source: https://gyglim.github.io/me/vsum/index.html

The UCF-Crime dataset is a large-scale dataset of 128 hours of videos. It consists of 1900 long and untrimmed real-world surveillance videos, with 13 realistic anomalies including Abuse, Arrest, Arson, Assault, Road Accident, Burglary, Explosion, Fighting, Robbery, Shooting, Stealing, Shoplifting, and Vandalism. These anomalies are selected because they have a significant impact on public safety. This dataset can be used for two tasks. First, general anomaly detection considering all anomalies in one group and all normal activities in another group. Second, for recognizing each of 13 anomalous activities. Source: Video Anomaly Dection Dataset

NTU RGB+D 120 is a large-scale dataset for RGB+D human action recognition, which is collected from 106 distinct subjects and contains more than 114 thousand video samples and 8 million frames. This dataset contains 120 different action classes including daily, mutual, and health-related activities. Source: NTU RGB+D 120: A Large-Scale Benchmark for 3D Human Activity Understanding

SEED-Bench consists of 19K multiple choice questions with accurate human annotations (~6 larger than existing benchmarks), which spans 12 evaluation dimensions including the comprehension of both the image and video modality.

The Freiburg-Berkeley Motion Segmentation Dataset (FBMS-59) is an extension of the BMS dataset with 33 additional video sequences. A total of 720 frames is annotated. It has pixel-accurate segmentation annotations of moving objects. FBMS-59 comes with a split into a training set and a test set. Source: https://lmb.informatik.uni-freiburg.de/resources/datasets/ Image Source: https://lmb.informatik.uni-freiburg.de/resources/datasets/

This dataset contains 118,081 short video clips extracted from 202 movies. Each video has a caption, either extracted from the movie script or from transcribed DVS (descriptive video services) for the visually impaired. The validation set contains 7408 clips and evaluation is performed on a test set of 1000 videos from movies disjoint from the training and val sets. Source: Use What You Have: Video Retrieval Using Representations From Collaborative Experts Image Source: https://sites.google.com/site/describingmovies/

The Georgia Tech Egocentric Activities (GTEA) dataset contains seven types of daily activities such as making sandwich, tea, or coffee. Each activity is performed by four different people, thus totally 28 videos. For each video, there are about 20 fine-grained action instances such as take bread, pour ketchup, in approximately one minute. Source: TricorNet: A Hybrid Temporal Convolutional and Recurrent Network for Video Action Segmentation Image Source: http://cbs.ic.gatech.edu/fpv/

VATEX is multilingual, large, linguistically complex, and diverse dataset in terms of both video and natural language descriptions. It has two tasks for video-and-language research: (1) Multilingual Video Captioning, aimed at describing a video in various languages with a compact unified captioning model, and (2) Video-guided Machine Translation, to translate a source language description into the target language using the video information as additional spatiotemporal context. Source: https://arxiv.org/pdf/1904.03493.pdf Image Source: https://arxiv.org/pdf/1904.03493.pdf

The 20BN-SOMETHING-SOMETHING dataset is a large collection of labeled video clips that show humans performing pre-defined basic actions with everyday objects. The dataset was created by a large number of crowd workers. It allows machine learning models to develop fine-grained understanding of basic actions that occur in the physical world. It contains 108,499 videos, with 86,017 in the training set, 11,522 in the validation set and 10,960 in the test set. There are 174 labels. ⚠️ Attention: This is the outdated V1 of the dataset. V2 is available here. Source: https://20bn.com/datasets/something-something/v1 Image Source: https://20bn.com/datasets/something-something/v1

AVA is a project that provides audiovisual annotations of video for improving our understanding of human activity. Each of the video clips has been exhaustively annotated by human annotators, and together they represent a rich variety of scenes, recording conditions, and expressions of human activity. There are annotations for: - Kinetics (AVA-Kinetics) - a crossover between AVA and Kinetics. In order to provide localized action labels on a wider variety of visual scenes, authors provide AVA action labels on videos from Kinetics-700, nearly doubling the number of total annotations, and increasing the number of unique videos by over 500x. - Actions (AvA Actions) - the AVA dataset densely annotates 80 atomic visual actions in 430 15-minute movie clips, where actions are localized in space and time, resulting in 1.62M action labels with multiple labels per human occurring frequently. - Spoken Activity (AVA ActiveSpeaker, AVA Speech). AVA ActiveSpeaker: associates speaking activity with a visible face, on the AVA v1.0 videos, resulting in 3.65 million frames labeled across ~39K face tracks. AVA Speech densely annotates audio-based speech activity in AVA v1.0 videos, and explicitly labels 3 background noise conditions, resulting in ~46K labeled segments spanning 45 hours of data. Image Source: https://www.researchgate.net/profile/Paolo_Napoletano/publication/309327222/figure/fig1/AS:419620126248965@1477056642346/Sample-images-from-the-Aesthetic-Visual-Analysis-AVA-database-sorted-by-their-aesthetic.png

The Penn Action Dataset contains 2326 video sequences of 15 different actions and human joint annotations for each sequence. Source: http://dreamdragon.github.io/PennAction/ Image Source: http://dreamdragon.github.io/PennAction/

SegTrack v2 is a video segmentation dataset with full pixel-level annotations on multiple objects at each frame within each video. Source: https://web.engr.oregonstate.edu/~lif/SegTrack2/dataset.html Image Source: https://www.researchgate.net/publication/325842926_Semantic_Video_Segmentation_A_Review_on_Recent_Approaches

The COIN dataset (a large-scale dataset for COmprehensive INstructional video analysis) consists of 11,827 videos related to 180 different tasks in 12 domains (e.g., vehicles, gadgets, etc.) related to our daily life. The videos are all collected from YouTube. The average length of a video is 2.36 minutes. Each video is labelled with 3.91 step segments, where each segment lasts 14.91 seconds on average. In total, the dataset contains videos of 476 hours, with 46,354 annotated segments. Source: COIN: A Large-scale Dataset for Comprehensive Instructional Video Analysis

As autonomous driving systems mature, motion forecasting has received increasing attention as a critical requirement for planning. Of particular importance are interactive situations such as merges, unprotected turns, etc., where predicting individual object motion is not sufficient. Joint predictions of multiple objects are required for effective route planning. There has been a critical need for high-quality motion data that is rich in both interactions and annotation to develop motion planning models. In this work, we introduce the most diverse interactive motion dataset to our knowledge, and provide specific labels for interacting objects suitable for developing joint prediction models. With over 100,000 scenes, each 20 seconds long at 10 Hz, our new dataset contains more than 570 hours of unique data over 1750 km of roadways. It was collected by mining for interesting interactions between vehicles, pedestrians, and cyclists across six cities within the United States. We use a high-accuracy 3D auto-labeling system to generate high quality 3D bounding boxes for each road agent, and provide corresponding high definition 3D maps for each scene. Furthermore, we introduce a new set of metrics that provides a comprehensive evaluation of both single agent and joint agent interaction motion forecasting models. Finally, we provide strong baseline models for individual-agent prediction and joint-prediction. We hope that this new large-scale interactive motion dataset will provide new opportunities for advancing motion forecasting models.

Kinetics-700 is a video dataset of 650,000 clips that covers 700 human action classes. The videos include human-object interactions such as playing instruments, as well as human-human interactions such as shaking hands and hugging. Each action class has at least 700 video clips. Each clip is annotated with an action class and lasts approximately 10 seconds.

The MovieQA dataset is a dataset for movie question answering. to evaluate automatic story comprehension from both video and text. The data set consists of almost 15,000 multiple choice question answers obtained from over 400 movies and features high semantic diversity. Each question comes with a set of five highly plausible answers; only one of which is correct. The questions can be answered using multiple sources of information: movie clips, plots, subtitles, and for a subset scripts and DVS. Source: Movie Question Answering: Remembering the Textual Cues for Layered Visual Contents Image Source: https://www.researchgate.net/figure/Examples-of-multiple-choice-QA-from-the-MovieQA-dataset-Each-question-has-5_fig2_321379716

Our task is to localize and provide a pixel-level mask of an object on all video frames given a language referring expression obtained either by looking at the first frame only or the full video. To validate our approach we employ two popular video object segmentation datasets, DAVIS16 [38] and DAVIS17 [42]. These two datasets introduce various challenges, containing videos with single or multiple salient objects, crowded scenes, similar looking instances, occlusions, camera view changes, fast motion, etc. DAVIS16 [38] consists of 30 training and 20 test videos of diverse object categories with all frames annotated with pixel-level accuracy. Note that in this dataset only a single object is annotated per video. For the multiple object video segmentation task we consider DAVIS17. Compared to DAVIS16, this is a more challenging dataset, with multiple objects annotated per video and more complex scenes with more distractors, occlusions, smaller objects, and fine structures. Overall, DAVIS17 consists of a training set with 60 videos, and a validation/test-dev/test-challenge set with 30 sequences each. As our goal is to segment objects in videos using language specifications, we augment all objects annotated with mask labels in DAVIS16 and DAVIS17 with non-ambiguous referring expressions. We follow the work of [34] and ask the annotator to provide a language description of the object, which has a mask annotation, by looking only at the first frame of the video. Then another annotator is given the first frame and the corresponding description, and asked to identify the referred object. If the annotator is unable to correctly identify the object, the description is corrected to remove ambiguity and to specify the object uniquely. We have collected two referring expressions per target object annotated by non-computer vision experts (Annotator 1, 2). However, by looking only at the 1st frame, the obtained referring expressions may potentially be invalid for an entire video. (We actually quantified that only∼ 15% of the collected descriptions become invalid over time and it does not affect strongly segmentation results as temporal consistency step helps to disambiguate some of such cases, see the supp. material for details.) Besides, in many applications, such as video editing or video-based advertisement, the user has access to a full video. Providing a language query which is valid for all frames might decrease the editing time and result in more coherent predictions. Thus, on DAVIS17 we asked the workers to provide a description of the object by looking at the full video. We have collected one expression of the full video type per target object. Future work may choose to use either setting. The average length for the first frame/full video expressions is 5.5/6.3 words. For DAVIS17 first frame annotations we notice that descriptions given by Annotator 1 are longer than the ones by Annotator 2 (6.4 vs. 4.6 words). We evaluate the effect of description length on the grounding performance in §5. Besides, the expressions relevant to a full video mention verbs more often than the first frame descriptions (44% vs. 25%). This is intuitive, as referring to an object which changes its appearance and position over time may require mentioning its actions. Adjectives are present in over 50% for all annotations. Most of them refer to colors (over 70%), shapes and sizes (7%) and spatial/ordering words (6% first frame vs. 13% full video expressions). The full video expressions also have a higher number of adverbs and prepositions, and overall are more complex than the ones provided for the first frame. Overall augmented DAVIS16/17 contains ∼ 1.2k referring expressions for more than 400 objects on 150 videos with ∼ 10k frames. We believe the collected data will be of interest to segmentation as well as vision and language communities, providing an opportunity to explore language as alternative input for video object segmentation.

FineGym is an action recognition dataset build on top of gymnasium videos. Compared to existing action recognition datasets, FineGym is distinguished in richness, quality, and diversity. In particular, it provides temporal annotations at both action and sub-action levels with a three-level semantic hierarchy. For example, a "balance beam" event will be annotated as a sequence of elementary sub-actions derived from five sets: "leap-jumphop", "beam-turns", "flight-salto", "flight-handspring", and "dismount", where the sub-action in each set will be further annotated with finely defined class labels. This new level of granularity presents significant challenges for action recognition, e.g. how to parse the temporal structures from a coherent action, and how to distinguish between subtly different action classes.

HACS is a dataset for human action recognition. It uses a taxonomy of 200 action classes, which is identical to that of the ActivityNet-v1.3 dataset. It has 504K videos retrieved from YouTube. Each one is strictly shorter than 4 minutes, and the average length is 2.6 minutes. A total of 1.5M clips of 2-second duration are sparsely sampled by methods based on both uniform randomness and consensus/disagreement of image classifiers. 0.6M and 0.9M clips are annotated as positive and negative samples, respectively. Authors split the collection into training, validation and testing sets of size 1.4M, 50K and 50K clips, which are sampled from 492K, 6K and 6K videos, respectively.

BABEL is a large dataset with language labels describing the actions being performed in mocap sequences. BABEL consists of action labels for about 43 hours of mocap sequences from AMASS. Action labels are at two levels of abstraction -- sequence labels describe the overall action in the sequence, and frame labels describe all actions in every frame of the sequence. Each frame label is precisely aligned with the duration of the corresponding action in the mocap sequence, and multiple actions can overlap. There are over 28k sequence labels, and 63k frame labels in BABEL, which belong to over 250 unique action categories. Labels from BABEL can be leveraged for tasks like action recognition, temporal action localization, motion synthesis, etc.

The PKU-MMD dataset is a large skeleton-based action detection dataset. It contains 1076 long untrimmed video sequences performed by 66 subjects in three camera views. 51 action categories are annotated, resulting almost 20,000 action instances and 5.4 million frames in total. Similar to NTU RGB+D, there are also two recommended evaluate protocols, i.e. cross-subject and cross-view. Source: Co-occurrence Feature Learning from Skeleton Data for Action Recognition and Detection with Hierarchical Aggregation Image Source: https://www.icst.pku.edu.cn/struct/Projects/PKUMMD.html

The UT-Kinect dataset is a dataset for action recognition from depth sequences. The videos were captured using a single stationary Kinect. There are 10 action types: walk, sit down, stand up, pick up, carry, throw, push, pull, wave hands, clap hands. There are 10 subjects, Each subject performs each actions twice. Three channels were recorded: RGB, depth and skeleton joint locations. The three channel are synchronized. The framerate is 30f/s. Source: https://cvrc.ece.utexas.edu/KinectDatasets/HOJ3D.html Image Source: https://cvrc.ece.utexas.edu/KinectDatasets/HOJ3D.html

This YouTube dataset is a sampling from thousands of User Generated Content (UGC) as uploaded to YouTube distributed under the Creative Commons license. This dataset was created in order to assist in the advancement of video compression and quality assessment research of UGC videos.

The great variations of videographic skills in videography, camera designs, compression and processing protocols, communication and bandwidth environments, and displays leads to an enormous variety of video impairments. Current no-reference (NR) video quality models are unable to handle this diversity of distortions. This is true in part because available video quality assessment databases contain very limited content, fixed resolutions, were captured using a small number of camera devices by a few videographers and have been subjected to a modest number of distortions. As such, these databases fail to adequately represent real world videos, which contain very different kinds of content obtained under highly diverse imaging conditions and are subject to authentic, complex and often commingled distortions that are difficult or impossible to simulate. As a result, NR video quality predictors tested on real-world video data often perform poorly. Towards advancing NR video quality prediction, we have constructed a large-scale video quality assessment database containing 585 videos of unique content , captured using 101 different devices (43 device models) by 80 different users with wide ranges of levels of complex, authentic distortions. We collected a large number of subjective video quality scores via crowdsourcing. A total of 4776 unique participants took part in the study, yielding more than 205000 opinion scores , resulting in an average of 240 recorded human opinions per video . This study is the largest video quality assessment study ever conducted along several key dimensions: number of unique contents, capture devices, distortion types and combinations of distortions, study participants, and recorded subjective scores.

WLASL is a large video dataset for Word-Level American Sign Language (ASL) recognition, which features 2,000 common different words in ASL. Source: https://github.com/dxli94/WLASL

The CAD-60 and CAD-120 data sets comprise of RGB-D video sequences of humans performing activities which are recording using the Microsoft Kinect sensor. Being able to detect human activities is important for making personal assistant robots useful in performing assistive tasks. The CAD dataset comprises twelve different activities (composed of several sub-activities) performed by four people in different environments, such as a kitchen, a living room, and office, etc. Source: https://www.re3data.org/repository/r3d100012216 Image Source: https://www.researchgate.net/figure/The-CAD-120-dataset-A-Examples-of-high-level-activities-from-the-dataset-B-A_fig3_335424041