Action Detection

Objective

The objective of the Action Detection lab is to experiment a particular action detection algorithm [1] with various template and search videos from the KTH dataset [2].

Software

A GUI is provided to allow the user to:
1. load a template video that contains the action of interest,
2. load a search video to detect the action of interest,
3. specify the number of detection results the user wishes to obtain in the search video, and
4. specify how similar the template video must be in the search video.

To access the necessary files for this lab, see directory /eecs/home/wildes/prism/4422/Lab5_ActionDetection

To start the GUI on Ubuntu,
1. Start MATLAB through the terminal
2. In MATLAB, change the working directory to the folder Lab5_ActionDetection
3. In MATLAB, run the GUI

Task

Execute the following example to detect a walking person
1. Load a video that contains the action of interest, e.g. ActionDetection/videos/person01_walking_d1_uncomp.mp4
2. Excise a template from the video, e.g. set the position from left and top to 40 and 0, resp.; set width and height to 40 and 50; set the frame to start and end at 14 and 44, resp.
3. Load the video to search the actions from, ActionDetection/videos/person03_walking_d1_uncomp.mp4.
4. Specify the maximum number of results you wish to detect, e.g. 10.
5. Specify the minimum similarity score, a real number between 0 and 1, the template must have to the detected regions in the search video, e.g. 0.50.
6. Click Detect & Localize to obtain the detection results.
7. Change the Display Options to Detected Results and press Play to visualize the detected results.

Repeat the same series of steps for different template actions (e.g. handclapping, handwaving, and walking) on various search videos (e.g. person01_handclapping_d1_uncomp.mp4, person03_handclapping_d4_uncomp.mp4, ...) with varying parameter settings (i.e. number of detects, threshold).

References

[1] K.G. Derpanis, M. Sizintsev, K. Cannons and R.P. Wildes, Action spotting and recognition based on a spatiotemporal orientation analysis, IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) 35, 2013.
[2] C. Schuldt, I. Laptev, and B. Caputo. Recognizing Human Action: A Local SVM Approach. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 32-36, 2004.