daking2014 / aslRecognition

aslRecognition

Performance on bcmnw (testing on person C):

• SVC(kernel=linear, C=1), color: 0.51
• SVC(kernel=linear, C=1), depth: 0.88
• SVC(kernel=linear, C=1), color and depth: 0.85

Parameters chosen by leave one person out cv:

• SVC(kernel=linear, C=0.001, best for color), color: 0.47
• SVC(kernel=linear, C=0.01, best for depth), depth: 0.89
• SVC(kernel=linear, C=0.001, best for color and depth), color and depth: 0.86
• SVC(kernel=rbf, C=10, gamma=0.001, best for color), color: 0.37
• SVC(kernel=rbf, C=10, gamma=0.001, best for depth), depth: 0.91
• SVC(kernel=rbf, C=100, gamma=0.0001, best for color and depth), color and depth: 0.88

Performance on overfeat_bcmnw (testing on person C):

Parameters chosen by leave one person out cv:

• SVC(kernel=linear, C=0.01, best for color), color: 0.67
• SVC(kernel=linear, C=0.01, best for depth), depth: 0.87
• SVC(kernel=linear, C=0.01, best for color and depth), color and depth: 0.84 (interestingly, the cv accuracies were much higher (0.92) than for the other settings)
• SVC(kernel=rbf, C=10, gamma=0.0001, best for color), color: 0.67
• SVC(kernel=rbf, C=10, gamma=0.0001, best for depth), depth: 0.88
• SVC(kernel=rbf, C=10, gamma=0.00001, best for color and depth), color and depth: 0.84

Performance with additional classifiers and features (bcmnw)

• KNN(k=3) depth (scaled 0.2): 0.86
• KNN(k=9) gabor: 0.84
• KNN(k=3) depth (0.2 scale) + gabor: 0.85
• SVC(kernel=linear, C=0.1) gabor: 0.88
• SVC(kernel=linear, C=0.1) depth (0.2 scale) + gabor: 0.95
• SVC(kernel=rbf, C=10, gamma=0.01) depth (0.2 scale) + gabor: 0.96
• LogReg(C=1.0,multinomial) depth (0.2 scale): 0.80
• LogReg(C=1.0,multinomial) gabor: 0.86
• LogReg(C=1.0,multinomial) gabor + depth (0.2 scale): 0.91
• RF('max_features': 0.28704608069377779, 'n_estimators': 19, 'max_depth': 8) overfeat: 0.76
• RF('max_features': 0.46912474635711165, 'n_estimators': 15, 'max_depth': 6) gabor: 0.76
• RF('max_features': 0.25839424397640276, 'n_estimators': 17, 'max_depth': 8) depth (0.2 scale): 0.53

Ensemble Performance

So far only found one ensemble that improves at all:

• LogRegOverfeat, SVMLinearDepth03Gabor, SVMRBFLinearDepth03Gabor: 0.963

Performance on all letters

• SVC(kernel=rbf, C = 100.0, gamma = 0.01) depth@0.3 gabor: 0.64
• SVC(kernel=linear, C = 10) depth@0.3 gabor: 0.65
• SVC(kernel=linear, C = 0.01) overfeat[depth]: 0.69
• SVC(kernel=rbf, C = 100, gamma = 0.0001) overfeat[depth]: 0.70
• Ensemble of
  • SVC(kernel=linear, C = 10) depth@0.3 gabor
  • SVC(kernel=linear, C = 0.01) overfeat[depth] had performance: 0.73
• Ensemble of
  • SVC(kernel=rbf, C = 100.0, gamma = 0.01) depth@0.3 gabor: 0.64
  • SVC(kernel=linear, C = 10) depth@0.3 gabor: 0.65
  • SVC(kernel=linear, C = 0.01) overfeat[depth]: 0.69
  • SVC(kernel=rbf, C = 100, gamma = 0.0001) overfeat[depth]: 0.70 had performance: 0.74