Hollywood 3D: A dataset for 3D action recognition in the wild

Hollywood 3D: A dataset for 3D action recognition in the wild

Posted by Rebecca Martin on Wed, 10/07/2013 - 10:56

A new dataset for “in the wild” action recognition has been released. The dataset is based on sequences taken from 3D movies, and comprises of 650 examples, across 14 action classes. There are plans to increase the size of the dataset further when additional 3D movies are available. http://personal.ee.surrey.ac.uk/Personal/S.Hadfield/hollywood3d.php

Each example in the dataset is well localized in time, and consists of a left viewpoint video, a right viewpoint video, and a reconstructed disparity video.

In addition to the dataset, source code is provided for 4D extensions of various natural action recognition techniques, including: Harris detector, Hessian Point detector, Dollars separable filter detector, Bag of visual words features and relative motion descriptors.