This Experiment was implemented using raw EEG signals on data set V of BCI competition III. Researchers: Shiliang Sun, E-mail: sunsl02@mails.tsinghua.edu.cn Changshui Zhang, E-mail: zcs@mail.tsinghua.edu.cn Jie Pan, E-mail: panj03@mails.tsinghua.edu.cn Affiliations: Shiliang Sun, Changshui Zhang: Department of Automation, Tsinghua University, Beijing, China, 100084. Jie Pan: Department of Biomedical Engineering, Tsinghua University, Beijing, China, 100084. Description: We first eliminated some electrodes with artifacts (eye movements and muscle movements). As a result, fifteen electrodes were retained, which are C3 Cz C4 CP1 CP2 P3 Pz P4 F3 Fz F4 FC1 FC2 PO3 PO4. We partitioned the continuous recordings to epochs of 1 second with 0.5 second overlapped. The recordings were downsampled to 128Hz, common average referenced and filtered with passbands 8-13Hz (for subjects 1 and 2) and 11-15Hz (for subject 3). Multiclass CSP was carried out to extract energy features for imaginary movement sources. A SVM classifier was trained with selected training sessions (those sessions with good generalization were retained). For subject 1, sessions 1 and 2 were finally used as training set. For subject 2, sessions 2 and 3 were finally used as training set. For subject 3, all the sessions 1, 2 and 3 were finally used as training set. The number of CSP sources were determined by the cross-validation results on training data. The test sessions were also partitioned to epoches of 1 second with 0.5 second overlapped. For example, if a test session lasts 240 seconds, then we can get 240*2-1 epoches. With the trained SVM classifier, the estimated labels for these test epoches were obtained and provided in Matlab (Version 6.5) data file 'EstimatedLabels.mat'. There are three variables in the file 'EstimatedLabels.mat', which are 'Subject1_EstimatedY', 'Subject2_EstimatedY', and 'Subject3_EstimatedY' corresponding to the estimated labels for the test sessions of subjects 1, 2, and 3 respectively.