1. MSR Action3D AS3:

99.14% Learn sparse dictionary+Basis normalization (

BasisMatrix=NormalizeMatrix(BasisMatrix); in Main_HistFeature_MSRAction3D.m) +direct projection rather than sparse coding (

ProjectCoeff=BasisMatrix*SingleVideo; 

in file Main_HistFeature_MSRAction3D.m)+Coefficient thresholding with thresholding value=0.01. 2.  I accidentally found that the performance was much better when the training subset for dictionary is different from the testing subset. For example, using AS1 as training set for dictionary and AS2 as training and testing set will lead much better performance. 3. Sparse dictionary is better than ICA dictioary. Orthogonal projection is better than sparse coding. 4. parameter setting: segment (f_s): 11 frames for CAD60 (MSRDaily: 23 frame; MSRAction: 13 frame); Num of Words (N_w): 400;  Expansion_Gamma=0.52; SVM_lambda=0.001; Coefficient_Threshold (0.02 for CAD60 and MSRDaily, 0.01 for MSRAction3D).

l0 dictionary+SparseFeature: CAD60 Dataset: Average Accuracy: 88.23; Average Precision: 80.95; Average Recall: 85.71 MSRAction3D: Average Accuracy: 85.59(AS1); 58.04(AS2);85.59(AS3);69.23(overall) MSRDaily Dataset: Average Accuracy: 47.50%

lo dictionary+DirectProjection: CAD60 Dataset: Average Accuracy: 94.12; Average Precision: 89.88; Average Recall: 92.86 MSRAction3D: Average Accuracy: 81.90(AS1); 83.04(AS2);97.30(AS3);86.81(overall) MSRDaily Dataset: Average Accuracy: 68.75% The above results show: sparse feature is not necessary for performance improvement and we know computing sparse feature is highly slow for sparse coding of  thousands of patches in testing phase. As what the paper " " says "over complete dictionary is not necessary for object recognition"