Affiliation: School of Mechanical Engineeing, Shanghai Jiao Tong University, 200240,China.
Name: Liu Guangquan, Huang Gan, and Zhu Xiangyang.
email: quanguangliu@sjtu.edu.cn

1. The EEG data were band-pass filtered between 8-30Hz by a 5-order IIR filter.  By this processing, EOG artifacts were also removed.
2. For each two of the four classes, CSP was applied to find the spatial filters W (resulting in 6 different W). Log-variances of the eight best components were used as features, resulting in a 6*8=48 dimensional feature vector for each trial at each time point. Time period used to train the CSP was determined by a ten-fold cross validation in the training data set. For test data set, signals from second 3 (when the imagery began) to current time were used for feature extraction. 
3. Fisher's LDA was used to reduce the dimension from 48 to 3.
4. A Bayesian classifier was used to classify the 3-dimensional features. For each test trial, only time period between second 3-6 was provided with estimated labels.
5. The 'Result_IIa.mat' file submitted were formed as follows: 'LabelEval(t,k,s)' contains the estimated label of the evaluation datasets, for time point t, trial k, and subject s. 'LabelTrain' contains the estimated label of training datasets. Labels of trials marked as artifacts were also provided.