iFRAME (Inhomogeneous Filters Random Field And Maximum Entropy)

Experiment 3.3: Learning from 100+ images with automatic alignment

Code and dataset

Case 1: Cat

Synthesized images by the learned models after alignment

Training images before alignment:

Training images after alignment:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=150; sizeTemplatey=150; GaborScaleList=[0.7]; DoGScaleList=[]; GaborScaleListLastIteration=[1.4, 1, 0.7, 0.5]; DoGScaleListLastIteration=[18.90, 13.36]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch= 300;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Alignment Parameters: flipOrNot = false; rotateShiftLimit=2; numberResolution=1; scaleStepSize=0.1; RatioDisplacementSUM2=0.25; numIterationEM=2;

Case 2: Deer

Synthesized images by the learned models before & after alignment

Training images before alignment:

Training images after alignment:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=150; sizeTemplatey=150; GaborScaleList=[0.7]; DoGScaleList=[]; GaborScaleListLastIteration=[1.4, 1, 0.7, 0.5]; DoGScaleListLastIteration=[18.90, 13.36]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch= 300;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Alignment Parameters: flipOrNot = false; rotateShiftLimit=2; numberResolution=1; scaleStepSize=0.1; RatioDisplacementSUM2=0.25; numIterationEM=2;

Case 3: Horse

Synthesized images by the learned models after alignment

Training images before alignment:

Training images after alignment:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=150; sizeTemplatey=200; GaborScaleList=[0.7]; DoGScaleList=[]; GaborScaleListLastIteration=[1.4, 1, 0.7, 0.5]; DoGScaleListLastIteration=[18.90, 13.36]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch= 300;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Alignment Parameters: flipOrNot = false; rotateShiftLimit=2; numberResolution=1; scaleStepSize=0.1; RatioDisplacementSUM2=0.25; numIterationEM=2;

Case 4: Wolf

Synthesized images by the learned models after alignment

Training images before alignment:

Training images after alignment:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=150; sizeTemplatey=150; GaborScaleList=[0.7]; DoGScaleList=[]; GaborScaleListLastIteration=[1.4, 1, 0.7, 0.5]; DoGScaleListLastIteration=[18.90, 13.36]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch= 300;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Alignment Parameters: flipOrNot = false; rotateShiftLimit=2; numberResolution=1; scaleStepSize=0.1; RatioDisplacementSUM2=0.25; numIterationEM=2;

Case 5: Duck

Synthesized images by the learned models after alignment

Training images before alignment:

Training images after alignment:

parameter setting:

General Parameters: nOrient = 16; sizeTemplatex=150; sizeTemplatey=150; GaborScaleList=[0.7]; DoGScaleList=[]; GaborScaleListLastIteration=[1.4, 1, 0.7, 0.5]; DoGScaleListLastIteration=[18.90, 13.36]; sigsq = 10; locationShiftLimit = 3; orientShiftLimit = 1; numSketch= 300;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Alignment Parameters: flipOrNot = true; rotateShiftLimit=2; numberResolution=1; scaleStepSize=0.1; RatioDisplacementSUM2=0.25; numIterationEM=2;