iFRAME (inhomogeneous Filters Random Field And Maximum Entropy)

Experiment 2: Learning Sparse FRAME by Generalized Gibbs Sampler and Generative Epsilon-boosting (multiple selection)

Code and dataset


Contents
Case 1 : Cat
Case 2 : Pigeon
Case 3 : Hummingbird
Case 4 : Tiger
Case 5 : Bike
Case 6 : Zebra
Case 7 : Leopard


Case 1: Cat

Training images:

synthesis by learning sparse-Frame model

Regular Version (with DoG)

(click here for movie of learning process)

Sketch template

Local normalization (without DoG)

(click here for movie of learning process)

Sketch template

parameters setting:

General Parameters: nOrient = 16; GaborScaleList=[1.4, 1, 0.7, 0.5]; DoGScaleList =[18.90, 13.36]; LocationShiftLimit=2; OrientShiftLimit=1; interval=5; #Wavelet=350;
Multiple Selection Parameters: nPartCol=5; nPartRow=5; part_sx=20; part_sy=20; sizeTemplatex=nPartRow*part_sx; sizeTemplatey=nPartCol*part_sy; gradient_threshold_scale=0.8;
Gibbs Sampler Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); 6x6 chains; sigsq=10; threshold_corrBB=0; lower_bound_rand = 0.001; upper_bound_rand = 0.999; c_val_list=-25:2:25
PS: If using local normalization, DoGScaleList =[]; localNormScaleFactor=2; thresholdFactor=0.01.


Case 2: Pigeon

Traning images

Synthesis by learning sparse-Frame model

Regular Version (with DoG)

(click here for movie of learning process)

Sketch template

Local normalization (without DoG)

(click here for movie of learning process)

Sketch template

parameters setting:

General Parameters: nOrient = 16; GaborScaleList=[1.4, 1, 0.7, 0.5]; DoGScaleList =[18.90, 13.36]; LocationShiftLimit=2; OrientShiftLimit=1; interval=5; #Wavelet=350;
Multiple Selection Parameters: nPartCol=5; nPartRow=5; part_sx=20; part_sy=20; sizeTemplatex=nPartRow*part_sx; sizeTemplatey=nPartCol*part_sy; gradient_threshold_scale=0.8;
Gibbs Sampler Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); 6x6 chains; sigsq=10; threshold_corrBB=0; lower_bound_rand = 0.001; upper_bound_rand = 0.999; c_val_list=-25:2:25
PS: If using local normalization, DoGScaleList =[]; localNormScaleFactor=2; thresholdFactor=0.01.


Case 3: Hummingbird

Training images

Synthesis by learning sparse-Frame model

Regular Version (with DoG)

(click here for movie of learning process)

Sketch template

Local normalization (without DoG)

(click here for movie of learning process)

Sketch template

parameters setting:

General Parameters: nOrient = 16; GaborScaleList=[1.4, 1, 0.7, 0.5]; DoGScaleList =[18.90, 13.36]; LocationShiftLimit=2; OrientShiftLimit=1; interval=5; #Wavelet=350;
Multiple Selection Parameters: nPartCol=5; nPartRow=5; part_sx=21; part_sy=25; sizeTemplatex=nPartRow*part_sx; sizeTemplatey=nPartCol*part_sy; gradient_threshold_scale=0.8;
Gibbs Sampler Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); 6x6 chains; sigsq=10; threshold_corrBB=0; lower_bound_rand = 0.001; upper_bound_rand = 0.999; c_val_list=-25:2:25
PS: If using local normalization, DoGScaleList =[]; localNormScaleFactor=2; thresholdFactor=0.01.


Case 4: Tiger

Training images

Synthesis by learning sparse-Frame model

Regular Version (with DoG)

(click here for movie of learning process)

Sketch template

Local normalization (without DoG)

(click here for movie of learning process)

Sketch template

parameters setting:

General Parameters: nOrient = 16; GaborScaleList=[1.4, 1, 0.7, 0.5]; DoGScaleList =[18.90, 13.36]; LocationShiftLimit=2; OrientShiftLimit=1; interval=5; #Wavelet=350;
Multiple Selection Parameters: nPartCol=5; nPartRow=5; part_sx=20; part_sy=20; sizeTemplatex=nPartRow*part_sx; sizeTemplatey=nPartCol*part_sy; gradient_threshold_scale=0.8;
Gibbs Sampler Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); 6x6 chains; sigsq=10; threshold_corrBB=0; lower_bound_rand = 0.001; upper_bound_rand = 0.999; c_val_list=-25:2:25
PS: If using local normalization, DoGScaleList =[]; localNormScaleFactor=2; thresholdFactor=0.01.


Case 5: Bike

Training images

Synthesis by learning sparse-Frame model

Regular Version (with DoG)

(click here for movie of learning process)

Sketch template

Local normalization (without DoG)

(click here for movie of learning process)

Sketch template

parameters setting:

General Parameters: nOrient = 16; GaborScaleList=[1.4, 1, 0.7, 0.5]; DoGScaleList =[18.90, 13.36]; LocationShiftLimit=2; OrientShiftLimit=1; interval=5; #Wavelet=350;
Multiple Selection Parameters: nPartCol=5; nPartRow=5; part_sx=20; part_sy=20; sizeTemplatex=nPartRow*part_sx; sizeTemplatey=nPartCol*part_sy; gradient_threshold_scale=0.8;
Gibbs Sampler Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); 6x6 chains; sigsq=10; threshold_corrBB=0; lower_bound_rand = 0.001; upper_bound_rand = 0.999; c_val_list=-25:2:25
PS: If using local normalization, DoGScaleList =[]; localNormScaleFactor=2; thresholdFactor=0.01.


Case 6: Zebra

Training images

Synthesis by learning sparse-Frame model

Regular Version (with DoG)

(click here for movie of learning process)

Sketch template

Local normalization (without DoG)

(click here for movie of learning process)

Sketch template

parameters setting:

General Parameters: nOrient = 16; GaborScaleList=[1.4, 1, 0.7, 0.5]; DoGScaleList =[18.90, 13.36]; LocationShiftLimit=2; OrientShiftLimit=1; interval=5; #Wavelet=350;
Multiple Selection Parameters: nPartCol=5; nPartRow=5; part_sx=21; part_sy=25; sizeTemplatex=nPartRow*part_sx; sizeTemplatey=nPartCol*part_sy; gradient_threshold_scale=0.8;
Gibbs Sampler Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); 6x6 chains; sigsq=10; threshold_corrBB=0; lower_bound_rand = 0.001; upper_bound_rand = 0.999; c_val_list=-25:2:25
PS: If using local normalization, DoGScaleList =[]; localNormScaleFactor=2; thresholdFactor=0.01.


Case 7: Leopard

Training images

Synthesis by learning sparse-Frame model

Regular Version (with DoG)

(click here for movie of learning process)

Sketch template

Local normalization (without DoG)

(click here for movie of learning process)

Sketch template

parameters setting:

General Parameters: nOrient = 16; GaborScaleList=[1.4, 1, 0.7, 0.5]; DoGScaleList =[18.90, 13.36]; LocationShiftLimit=2; OrientShiftLimit=1; interval=5; #Wavelet=350;
Multiple Selection Parameters: nPartCol=5; nPartRow=5; part_sx=20; part_sy=26; sizeTemplatex=nPartRow*part_sx; sizeTemplatey=nPartCol*part_sy; gradient_threshold_scale=0.8;
Gibbs Sampler Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); 6x6 chains; sigsq=10; threshold_corrBB=0; lower_bound_rand = 0.001; upper_bound_rand = 0.999; c_val_list=-25:2:25
PS: If using local normalization, DoGScaleList =[]; localNormScaleFactor=2; thresholdFactor=0.01.