特征描述子surf，hog，光流

Action Snippets V1.0 ==================== Code written by Konrad Schindler, using components provided by Peter Kovesi and Anton Schweighofer. Research by Konrad Schindler and L. van Gool. Quick start ----------- 1) uncompress the archive with TAR or WINZIP. 2) change to directory snippets/code 3) open MATLAB 4) run the feature extraction, using the provided data ad control file: > features('../data/weizmann.txt'); This takes a while. Features are written to file, so you need not do this again, unless you change the parameters of feature extraction. 5) run the feature pooling: > pooling('../data/weizmann.txt'); Again, the pooled features are written to file, so unless you change the pooling size, you need not do this again. 6) test the classification: > crossvalidation('../data/weizmann.txt'); This takes a LONG while, because the PCA-basis is recomputed for each fold. The PCA-bases are written to file, so that you can later load them from file, as long as you are using the same snippet length - change the variable <Globals.features_tofile>. Overview -------- Welcome to this software release, which implements the functionality reported in [1]: - extracting features from action videos - training a classifier for the actions - testing the classifier with crossvalidation The release includes the complete Matlab source code. It is recommend to use the code only with Matlab 8 or higher. In addition to the code, this release also includes the cropped frames of the WEIZMANN dataset [2]. The crops were generated automatically, using the author's silhouette annotations. Parameter Settings ------------------ All parameters are set directly in the top-level files. Look for variables names <Globals.XXX>. Variable names and code comments should be self-explanatory, if you read the paper [1]. Using your dataset ------------------ To use other data than the provided Weizmann frames: - split the video into single files; all frames must be the same size, and all clips must have the same number of frames. The frames from one clip must follow the naming convention <name.3-digit-framenumber.extension>. - prepare a control file, using the template <weizmann.txt>. Class labels are integer numbers starting from 1. - run the three steps of the program with your controlfile. Support ------- For any query/suggestion/complaint or simply to say you like/use this software, just drop me an email: [email protected] I wish you a good time using this software, Konrad References ---------- [1] K. Schindler and L. van Gool. Action snippets: how many frames does human action recognition require? CVPR 2009. [2] L. Gorelick, M. Blank, E. Shechtman, M. Irani and R. Basri. Actions as space-time shapes. ICCV 2005.