iFRAME (inhomogeneous Filters Random Field And Maximum Entropy)

Experiment 5.2: Unsupervised Learning of Codebooks (local normalization)

Code and dataset

new results

Case 1: Grape

Parameters setting

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[0.7]; DoGScaleList=[]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch = 37;  isGlobalNormalization = true; isLocalNormalize = true; resizeFactor = 1.2;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =40; 12x12 chains;
Coodebook Parameters: rotateShiftLimit = 16; allResolution = [0.8, 1, 1.2]; #EM iteration = 15; numCluster = 3; maxNumClusterMember = 1000; locationPerturbationFraction = 0.4;

Case 2: Beehive

training image

Parameters setting

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[0.7]; DoGScaleList=[]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch = 30;  isGlobalNormalization = true; isLocalNormalize = true; resizeFactor = 1.2;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =40; 12x12 chains;
Coodebook Parameters: rotateShiftLimit = 16; allResolution = [0.8, 1, 1.2]; #EM iteration = 15; numCluster = 1; maxNumClusterMember = 1000; locationPerturbationFraction = 0.4;

Case 3: Peanut

Parameters setting

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[0.7]; DoGScaleList=[]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch = 30;  isGlobalNormalization = true; isLocalNormalize = true; resizeFactor = 1.4;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =40; 12x12 chains;
Coodebook Parameters: rotateShiftLimit = 16; allResolution = [0.8, 1, 1.2]; #EM iteration = 15; numCluster = 1; maxNumClusterMember = 1000; locationPerturbationFraction = 0.4;

Case 4: Cat

Parameters setting

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[0.7]; DoGScaleList=[]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch = 40;  isGlobalNormalization = true; isLocalNormalize = true; resizeFactor = [160, 160];
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =40; 12x12 chains;
Coodebook Parameters: rotateShiftLimit = 2; allResolution = [0.8, 1, 1.2]; #EM iteration = 15; numCluster = 8; maxNumClusterMember = 1000; locationPerturbationFraction = 0.4;

Case 5: Lotus

Parameters setting

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[0.7]; DoGScaleList=[]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch = 30;  isGlobalNormalization = true; isLocalNormalize = true; resizeFactor = 1;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =40; 12x12 chains;
Coodebook Parameters: rotateShiftLimit = 16; allResolution = [0.8, 1, 1.2]; #EM iteration = 15; numCluster = 2; maxNumClusterMember = 1000; locationPerturbationFraction = 0.4;