MSR Action3D
20 action types, 10 subjects, each subject performs each action 2 or 3 times. There are 567 depth map sequences in total. The resolution is 640x240. The data was recorded with a depth sensor similar to the Kinect device. The dataset is described in the following paper.
Action Recognition Based on A Bag of 3D Points, Wanqing Li, Zhengyou Zhang, Zicheng Liu, IEEE International Workshop on CVPR for Human Communicative Behavior Analysis (in conjunction with CVPR2010), San Francisco, CA, June, 2010.
Code to load and display depth maps (Load MSRAction3D_depth.zip) is provided by Josue Rocha Lima <mailto::mjosuerocha@me.com>
Better classification results are reported in the following paper:
Mining Actionlet Ensemble for Action Recognition with Depth Cameras, Jiang Wang, Zicheng Liu, Ying Wu, Junsong Yuan, IEEE Conference on Computer Vision and Pattern Recognition (CVPR2012), Providence, Rhode Island, June 16-21, 2012.
Note that there is an error in the paper on the number of samples being used for the experiment. The number 402 in the paper is not correct. The correct number is 557. Out of the original 567 sequences in MSR Action3D Dataset, 10 sequences are not used in this paper's experiment because the skeletons are either missing or too erroneous.
Sample code to load MSR Action3D Dataset (drawskt.zip)
Skeleton Data in screen coordinates (MSRAction3DSkeleton (20joints).rar) (Thanks to Yi Wen Wan, University of North Texas, for data cleaning and conversion). There is a skeleton sequence file for each depth sequence in the Action3D dataset. A skeleton has 20 joint positions (see the image for illustrations of the joint positions). Four real numbers are stored for each joint: u, v, d, c where (u,v) are screen coordinates, d is the depth value, and c is the confidence score. If a depth sequence has n frames, then the number of real numbers stored in the corresponding skeleton file is equal to: n*20*4. Click here for MATLAB code to visualize the skeleton motions (The code is provided by Antonio Vieira from Federal University of Minas Gerais).
This diagram shows the correspondence between the 20 points in the skeleton data and the joints (Thanks to Yu Zhong from AIT, BAE Systems for providing this diagram).
Skeleton Data in real world coordinates (MSRAction3DSkeletonREal3D.rar) (Thanks to Ferda Ofli, UC Berkeley, for processing the data).