Data

Video Stream Retrieval Splits

When reporting results on these data set splits, please reference our BMVC submission.

The FCVS dataset consists of 25 videos per class of the FCVID dataset. These are split along class lines, and are also concatenated into longer videos (FCVS-L) for experiments on no-example live stream retrieval. The videos and the splits are below:

• FCVS Subset Videos
• FCVS Train Classes
• FCVS Test Classes
• FCVS-L Train Videos
• FCVS-L Test Classes

The splits used on the ActivityNet 1.0 data set are below, as well as the video names of the concatenated versions:

• AN Train Classes
• AN Test Classes
• AN-L Train Videos
• AN-L Test Videos

1M Micro-Blog Images (MBI-1M)

The MBI-1M dataset is a new collection of 1 million images from Twitter, along with accompanying tweets and metadata. The dataset was constructed by selecting 1 million of the image-containing tweets in the data from February and March 2013 of the Trec 2013 microblog track. Updated re-tweet and favorite counts were collected in Autumn 2014, so as to be useful for popularity prediction.

The images and data made available here are intended for research use, and we claim no copyright or responsibility over their content. Please refer to the Twitter Terms of Service for any questions regarding rights ownership.

When using this dataset, please cite our ICMR conference paper on popularity prediction.

• Images (63 GB), zipped archive
• Annotations & Metadata (3 GB), JSON formatted
• Train/Val/Test splits (28 MB), plaintext

VSO Popularity Annotations

The Visual Sentiment Ontology is a large dataset of 930k images from Flickr with annotations of adjective-noun pairs. For our work on predicting image popularity, we collected view and comment counts for 891,297 of these images. These values are available here:

• View Counts
• Comment Counts