PASCAL VOC 2010 Action Classification
Below is an implementation of the action classification algorithm described in the paper:
Action Recognition Using a Distributed Representation of Pose and Appearance,
Subhransu Maji, Lubomir Bourdev and Jitendra Malik,
In Proceedings, CVPR 2011, Colorado Springs.
The 2D keypoint annotations for the people in the PASCAL VOC 2010 action dataset can be downloaded here (VOC10action-annotations.tar.gz). These were collected on Amazon Mechanical Turk and were used to train action specific poselets as described in the paper. The data is in the H3D format.
For the 3D pose software go here: http://www.cs.berkeley.edu/~smaji/projects/3dpose
The code is built on top of the poselets framework available here.
Download and Instructions
Here is the stand alone code for action classification (action-classification-release1.0.tar.gz). The input is an image and a list of bounding boxes for each person. The output is the scores for each action class.
NOTE : The model is based on poselet activation vector only (Table 1, PAV in the paper).
- The code includes the poselet detection module. Make sure you can run it first by running the demo.m inside the poselet_detection folder. See here for more details.
- Run the demoAction.m file which estimates the action classes based on the poselet activation vector.
Below are the outputs of the demo:
Each image shows a person with the bounding box, and the top 3 actions with their scores in the title.
Last updated: March 26, 2011