Sparse iFRAME (Sparse Inhomogeneous Filters Random Field And Maximum Entropy)

Experiment 3.2: Objects Detection by Deformable Sparse FRAME Model (sqrt_version)

(Geometric transformation of template: allow shift of template in location, orientation and scale as well as left-right flip in detection)

Code and dataset

Case 1: Lion

(1) Filters selection by shared matching pursuit:

learned templates (Layout of Gabors)

(i) original (ii) deformed templates (iii) reconstruced (iv) residual

(2) Learning sparse-Frame model (by 6x6 multiple chains):

projected synthesized image (single chain)

projected synthesized image (multiple chains)

(3) Object detection by geometric transformation of template

(4) More testing images:

(5) Cropped objects:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[1.2, 0.7]; sigsq = 10; locationShiftLimit = 4; orientShiftLimit = 1; numSketch = 300;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Detection Parameters: flipOrNot = false; rotateShiftLimit=4; numberResolution=17; scaleStepSize=0.1; initializedScaleFactor=150.


Case 2: Tiger

(1) Filters selection by shared matching pursuit:

learned templates (Layout of Gabors)

(i) original (ii) deformed templates (iii) reconstruced (iv) residual

(2) Learning sparse-Frame model (by 6x6 multiple chains):

projected synthesized image (single chain)

projected synthesized image (multiple chains)

(3) Object detection by geometric transformation of template

(4) More testing images:

(5) Cropped objects:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[1.2, 0.7]; sigsq = 10; locationShiftLimit = 4; orientShiftLimit = 1; numSketch = 250;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Detection Parameters: flipOrNot = false; rotateShiftLimit=4; numberResolution=17; scaleStepSize=0.1; initializedScaleFactor=170.


Case 3: Wolf

(1) Filters selection by shared matching pursuit:

learned templates (Layout of Gabors)

(i) original (ii) deformed templates (iii) reconstruced (iv) residual

(2) Learning sparse-Frame model (by 6x6 multiple chains):

projected synthesized image (single chain)

projected synthesized image (multiple chains)

(3) Object detection by geometric transformation of template

(4) More testing images:

(5) Cropped objects:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[1.2, 0.7]; sigsq = 10; locationShiftLimit = 4; orientShiftLimit = 1; numSketch = 400;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Detection Parameters: flipOrNot = false; rotateShiftLimit=3; numberResolution=20; scaleStepSize=0.1; initializedScaleFactor=200.


Case 4: Cat

(1) Filters selection by shared matching pursuit:

learned templates (Layout of Gabors)

(i) original (ii) deformed templates (iii) reconstruced (iv) residual

(2) Learning sparse-Frame model (by 6x6 multiple chains):

projected synthesized image (single chain)

projected synthesized image (multiple chains)

(3) Object detection by geometric transformation of template

(4) More testing images:

(5) Cropped objects:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=100; sizeTemplatey=100; GaborScaleList=[1.2, 0.7]; sigsq = 10; locationShiftLimit = 4; orientShiftLimit = 1; numSketch = 250;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Detection Parameters: flipOrNot = false; rotateShiftLimit=4; numberResolution=17; scaleStepSize=0.1; initializedScaleFactor=200.


Case 5: Deer

(1) Filters selection by shared matching pursuit:

learned templates (Layout of Gabors)

(i) original (ii) deformed templates (iii) reconstruced (iv) residual

(2) Learning sparse-Frame model (by 6x6 multiple chains):

projected synthesized image (single chain)

projected synthesized image (multiple chains)

(3) Object detection by geometric transformation of template

(4) More testing images:

(5) Cropped objects:

parameters setting:

General Parameters: nOrient = 16; sizeTemplatex=120; sizeTemplatey=122; GaborScaleList=[1.4, 0.7]; sigsq = 10; locationShiftLimit = 6; orientShiftLimit = 1; numSketch = 270;
HMC Parameters: lambdaLearningRate = 0.1/sqrt(sigsq); epsilon = 0.03; L = 10; nIteraton =80;
Detection Parameters: flipOrNot = false; rotateShiftLimit=3; numberResolution=20; scaleStepSize=0.1; initializedScaleFactor=120.