Experiment Scripts

mainExperiment.m

@@ -0,0 +1,39 @@
+function [perfAblations] = mainExperiment()
+%MAINEXPERIMENT Summary of this function goes here
+%   Detailed explanation goes here
+globals;
+params = getParams();
+classInds = [6 14 12 17]; %bus, motorbike, dog, sheep
+nClasses = length(classInds);
+usePascalViews = [0 0 1 1]; %evaluation using pascal3d or pascal voc labels
+
+%% Run SCT vs GC Experiment
+perfAblations = zeros(5,nClasses);
+
+params.features = 'vggJoint16';
+perfAblations(1,:) = mainViewpoint(classInds, usePascalViews); %% prints accuracy for bus, motorbike, dog, sheep respectively
+
+params.features = 'vggCommon16';
+perfAblations(2,:) = mainViewpoint(classInds, usePascalViews); %% prints accuracy for bus, motorbike, dog, sheep respectively
+
+%% Run experiment for various similarity features
+
+params.similarityFeatName =  'vggConv5';
+for c = 1:nClasses
+    [~,optAcc] = optimizePredictions(pascalIndexClass(classInds(c),'pascal'),0,usePascalViews(c),1,0,0);
+    perfAblations(3,c) = optAcc;
+end
+
+params.spatialNormSmoothing = 0;
+for c = 1:nClasses
+    [~,optAcc] = optimizePredictions(pascalIndexClass(classInds(c),'pascal'),0,usePascalViews(c),1,0,0);
+    perfAblations(4,c) = optAcc;
+end
+
+params.similarityFeatName =  'vggFc7';
+for c = 1:nClasses
+    [~,optAcc] = optimizePredictions(pascalIndexClass(classInds(c),'pascal'),0,usePascalViews(c),1,0,0);
+    perfAblations(5,c) = optAcc;
+end
+
+end

mainPredict.m

@@ -3,14 +3,18 @@
 %   Detailed explanation goes here
 
 %% predictions for SCT
-generatePoseFeatures('vggJoint16','vggJoint16',224,[6 12 14 17],0,0,[7 9 2 10]);
+generatePoseFeatures('vggJoint16','vggJoint16',224,[6 12 14 17],2,0,[7 8 2 10]);
 caffe.reset_all();
 
 %% predictions for GC
-generatePoseFeatures('vggCommon16','vggCommon16',224,[6 12 14 17],0,0);
+generatePoseFeatures('vggCommon16','vggCommon16',224,[6 12 14 17],2,0);
 caffe.reset_all();
 
 %% add VGG conv5 feature computation here
+generatePoseFeatures('vggConv5','vgg',224,[6 12 14 17],2,0);
+caffe.reset_all();
 
+generatePoseFeatures('vggFc7','vgg',224,[6 12 14 17],2,0);
+caffe.reset_all();
 
 end

mainViewpoint.m

@@ -0,0 +1,24 @@
+function errors = mainViewpoint(classInds, usePascalViews)
+    %% Define Classes
+    classes = {'aeroplane','bicycle','bird','boat','bottle','bus','car','cat','chair','cow','diningtable','dog','horse','motorbike','person','plant','sheep','sofa','train','tvmonitor'};
+    %classInds = [1 2 4 5 6 7 9 11 14 18 19 20];
+    %classInds = [1 2 4 5 6 7 9];
+    %classInds = [1 2 6 7 9 14];
+    %classInds = [1 2 4 5 6 7 9 11 14 18 19 20];
+    %classInds = [1 2 4 5 6 7 9 14 18 19 20];
+    %classInds = [12 17];
+
+    numClasses = size(classInds,2);
+
+    %% Iterate over pose predictions
+    errors = zeros(numClasses,1);
+    medErrors = zeros(numClasses,1);
+    for c = 1:numClasses
+        class = classes{classInds(c)};
+        disp(class);
+        [err,medErr] = regressToPose(class, usePascalViews(c));
+        errors(c,:)=err;
+        medErrors(c,:) = medErr;
+    end
+    prettyPrintResults(errors,medErrors);
+end

optimization/optimizePredictions.m

@@ -1,4 +1,4 @@
-function [preds] = optimizePredictions(class,useSaved,evalPascalViews,useMirror,useSoftAssignment,azimuthOnly)
+function [preds,testAccuracyOpt] = optimizePredictions(class,useSaved,evalPascalViews,useMirror,useSoftAssignment,azimuthOnly)
 %OPTIMIZEPREDICTIONS Summary of this function goes here
 %   Detailed explanation goes here
 % useSaved is 0/1 : 1 means it uses
@@ -59,21 +59,22 @@
     end
 end
 %keyboard;
-similarityFeatName = params.similarityFeatName;'vggConv5';
+similarityFeatName = params.similarityFeatName; %'vggConv5';
 spatialFeat = ~isempty(strfind(similarityFeatName,'Pool')) || ~isempty(strfind(similarityFeatName,'Conv'));
 spatialNormSmoothing = params.spatialNormSmoothing;
 
 %% set-up optimization
 formulationDir = fullfile(cachedir,['optimizationInit' params.vpsDataset],params.features);
 mkdirOptional(formulationDir);
 if(useSaved && exist(fullfile(formulationDir,[class '.mat']),'file'))
-    load(fullfile(formulationDir,[class '.mat']))
+    load(fullfile(formulationDir,[class '.mat']));
 else
     fg = fspecial('gaussian',[5 5],1);
-    load(fullfile(cachedir,['rcnnPredsVps' params.vpsDataset],similarityFeatName,class));
+    var = load(fullfile(cachedir,['rcnnPredsVps' params.vpsDataset],similarityFeatName,class));
     %load(fullfile(cachedir,['rcnnPredsVps' params.vpsDataset],'vggSimilarityCommon16',class));
-    feat = feat(goodInds);
+    feat = var.featStruct{1};feat = feat(goodInds);
     if(useMirror)
+        featMirror = var.featStructMirror{1};
         featMirror = featMirror(goodInds);
         feat = vertcat(feat,featMirror);
     end
@@ -89,6 +90,7 @@
     else
         for i=1:length(feat)
             feat{i} = sigmoid(feat{i});
+            feat{i} = (feat{i}(:))';
         end
     end
 
@@ -162,8 +164,8 @@
     end
 
     [testErrsOpt] = evaluatePredictionError({preds},testLabels,encoding,0);
-    testAccuracyOpt = sum(testErrsOpt<=30)/numel(testErrsOpt)
-    testMedErrorOpt = median(testErrsOpt)
+    testAccuracyOpt = sum(testErrsOpt<=30)/numel(testErrsOpt);
+    testMedErrorOpt = median(testErrsOpt);
     disp(numFlips);
 end
 
@@ -183,9 +185,9 @@
 %% eval left/right/frontal
 if(evalPascalViews)
     [accLabels,~,isCorrectLabels] = evaluatePascalViews(testLabels(:,3),data.test.views);
-    accLabels
+    %accLabels
     [accOpt,isGoodOpt,isCorrectOpt] = evaluatePascalViews(preds(:,3),data.test.views);
-    accOpt    
+    testAccuracyOpt = accOpt;    
     evaluatePascalViews(testPredsAec{1}(:,3),data.test.views)
 end
 

prettyPrintResults.m

@@ -7,9 +7,9 @@
 numClasses = size(errors,1);
 for col = 1:numCol
     for c = 1:numClasses
-        fprintf('%2.2f (%2.2f) \n',errors(c,col),medErrors(c,col));
+        fprintf('%2.2f\n',errors(c,col));
     end
-    fprintf('%2.2f (%2.2f) \n\n\n',mean(errors(:,col)),mean(medErrors(:,col)));
+    %fprintf('%2.2f (%2.2f) \n\n\n',mean(errors(:,col)),mean(medErrors(:,col)));
 end
 
 end

regressToPose.m

@@ -1,4 +1,4 @@
-function [testErrors,testMedErrors,testErrs,testData,testPreds,testLabels] = regressToPose(class)
+function [testErrors,testMedErrors,testErrs,testData,testPreds,testLabels] = regressToPose(class, evalPascalViews)
 %[testErrors,testMedErrors] = regressToPose(class)
 %   uses the training/val/test sets specified in parameters and
 % regresses to pose and returns error
@@ -14,7 +14,7 @@
 
 %data = load(fullfile(cachedir,'splitSets',class));
 data = load(fullfile(cachedir,'evalSets',class));
-[trainLabels,valLabels,testLabels,trainFeats,valFeats,testFeats] = generateEvalSetData(data);
+[~,~,testLabels,~,~,testFeats] = generateEvalSetData(data);
 
 %% TESTING
 switch params.optMethod        
@@ -36,6 +36,9 @@
 
 %keyboard;
 testErrs = evaluatePredictionError(testPreds,testLabels,encoding,0);
+if(evalPascalViews)
+    [accPascalViews] = evaluatePascalViews(testPreds(:,3),data.test.views);
+end
 %[valErrs,bestValPred] = evaluatePredictionError(valPreds,valLabels,encoding,0);
 %[trainErrs,bestTrainPred] = evaluatePredictionError(trainPred,trainLabels,encoding,0);
 
@@ -44,9 +47,6 @@
 testErrors = [];testMedErrors=[];
 %testErrs = [valErrs;testErrs];
 
-testErrors(1) = mean(testErrors);
-testMedErrors(1) = median(testErrors);
-
 % plot(sort(valErrs),[1:length(valErrs)]./length(valErrs),'r');hold on;
 % plot([0 180],[0 1],'k');
 % grid on;xlim([0 180]);ylim([0 1]);
@@ -56,7 +56,12 @@
 
 testErrors(1) = sum(testErrs<=30)/numel(testErrs);
 testMedErrors(1) = median(testErrs);
+if(evalPascalViews)
+    [accPascalViews] = evaluatePascalViews(testPreds(:,3),data.test.views);
+    testErrors(1) = accPascalViews;
+end
 testData = data.test;
+
 %[errSort,IDX] = sort(testErrs,'ascend');
 %plot(errSort,[1:length(errSort)]/length(errSort));pause();close all;
 %visualizePredictions(class,testPreds{1},data.test,encoding,testErrs,'image');