| Home | Publications | Teaching | Resources | Contacts |
|---|
The MSRC-12 dataset is annotated with the action point of each action instance. This annotation is useful for evaluating action detection techniques, which aims at locating an action instance in a video sequence. For action recognition, it could be useful to know the boundaries, i.e. the onset and offset of each action instance. This is what we provide here.
You can download our annotation, along with our annotation tool, from here. The annotation tool assumes that you already have the original dataset and the code provided with it (a slightly modified copy is available with the annotation tool).
We assume that you already have the original dataset and the code associated with it. The annotation is provided in the form of a list of text files, each file is associated with a sequence file from the original dataset. For the sequence file file.csv from the original dataset, our annotation is in the file file.sep. Each *.sep file is a text file. Each line of the text file contains the starting and ending frame indices of an action instance in the associated sequence file. The frame indices are 1-based and cover only frames returned using the load_file.m function, which is provided with the code of the original dataset.
We hope you find this annotation useful. In case you use it, please cite the following paper.
Mohamed E. Hussein, Marwan Torki, Mohammad A. Gowayyed, and Motaz El-Saban, "Human Action Recognition Using a Temporal Hierarchy of Covariance Descriptors on 3D Joint Locations," International Joint Conference on Artificial Intelligence (IJCAI), Beijing, China, August 2013. [pdf][bibtex][code][ERRATA]