• Sep 4th, 2019: Released the MediaPipe YouTube-8M feature extractor which extracts both visual and audio features.
• Jun 27th, 2019: Released the YouTube-8M Segments dataset.
• May 14th, 2018: Released an update to the dataset, with improved quality machine-generated labels, and reduced size / higher-quality video dataset. (YouTube-8M 2018).

The YouTube-8M Segments dataset is an extension of the YouTube-8M dataset with human-verified segment annotations. In addition to annotating videos, we would like to temporally localize the entities in the videos, i.e., find out when the entities occur.

We collected human-verified labels on about 237K segments on 1000 classes from the validation set of the YouTube-8M dataset. Each video will again come with time-localized frame-level features so classifier predictions can be made at segment-level granularity. We encourage researchers to leverage the large amount of noisy video-level labels in the training set to train models for temporal localization.

We are organizing a Kaggle Challenge and The 3rd Workshop on YouTube-8M Large-Scale Video Understanding at ICCV 2019.

237K Human-verified Segment Labels

1000 Classes

5.0 Avg. Segments / Video

Dataset Vocabulary

The vocabulary of the segment-level dataset is a subset of the YouTube-8M dataset (2018 version) vocabulary. We exclude the entities that are not temporally localizable like movies or TV series, which usually occurs in the whole video.

The following figure shows the distribution of the ratings in the YouTube-8M Segments dataset. Each class contains up to 250 human ratings (indicated by the grey bar in the background). The number of positives (indicated by the red bar) varies between classes.

YouTube-8M is a large-scale labeled video dataset that consists of millions of YouTube video IDs, with high-quality machine-generated annotations from a diverse vocabulary of 3,800+ visual entities. It comes with precomputed audio-visual features from billions of frames and audio segments, designed to fit on a single hard disk. This makes it possible to train a strong baseline model on this dataset in less than a day on a single GPU! At the same time, the dataset's scale and diversity can enable deep exploration of complex audio-visual models that can take weeks to train even in a distributed fashion.

Our goal is to accelerate research on large-scale video understanding, representation learning, noisy data modeling, transfer learning, and domain adaptation approaches for video. More details about the dataset and initial experiments can be found in our technical report and in previous workshop pages (2018, 2017). Some statistics from the latest version of the dataset are included below.

6.1 Million Video IDs

350,000 Hours of Video

Inception-V3 image annotation model, trained on ImageNet. The audio features were extracted using a VGG-inspired acoustic model described in Hershey et. al. on a preliminary version of YouTube-8M. Both the visual and audio features were PCA-ed and quantized to fit on a single hard disk. The combined set of all features are less than 2TB in size."> 2.6 Billion Audio/Visual Features

Knowledge Graph entities, including both coarse and fine-grained entities, which have been semi-automatically curated and manually verified by 3 raters to be visually recognizable. Each entity has at least 200 corresponding video examples, with an average of 3552 training videos per entity. The three most popular entities are Game, Video Game, and Vehicle, respectively, with 788288, 539945, and 415890 training examples, respectively. The least frequent are Cylinder and Mortar, with 123 and 127 training videos, respectively. The entities are grouped into 24 high-level verticals, with the most frequent vertical being Arts & Entertainment (3.3M training videos) and the least frequent being Finance (6K training videos)."> 3862 Classes

3.0 Avg. Labels / Video

The (multiple) labels per video are Knowledge Graph entities, organized into 24 top-level verticals. Each entity represents a semantic topic that is visually recognizable in video, and the video labels reflect the main topics of each video.

You can download a CSV file (2017 version CSV, deprecated) of our vocabulary. The first field in the file corresponds to each label's index in the dataset files, with the first label corresponding to index 0. The CSV file contains the following columns:

Index,TrainVideoCount,KnowledgeGraphId,Name,WikiUrl, Vertical1,Vertical2,Vertical3,WikiDescription

The entity frequencies are plotted below in log-log scale, which shows a Zipf-like distribution:


In addition, we show histograms with the number of entities and number of training videos in each top-level vertical:

This dataset is brought to you from the Video Understanding group within Google Research. More about us.
If you want to stay up-to-date about this dataset, please subscribe to our Google Group: youtube8m-users. The group should be used for discussions about the dataset and the starter code.