COLOR-ATTRIBUTES-RELATED IMAGE RETRIEVAL WEEK 4 Student: Kylie Gorman Mentor: Yang Zhang GMM AND FISHER VECTOR CODE STEP ONE: CONCATENATE THE FEATURE MATRICES OF EACH IMAGE STEP TWO: APPLY GMM FUNCTION Generate mean, covariance, and prior mode probabilities Mean Covariance Prior Mode Probabilities STEP THREE: CREATE FISHER VECTORS encoding = vl_fisher(new', means, covariances, priors); REPEAT PROCESS FINAL STEPS Calculate feature matrix of each image, isolating the object first Concatenate matrices Apply PCA function to preprocess data Multiply each individual feature matrix by result Concatenate output into 1 matrix Apply GMM function and obtain mean, covariance, and prior mode probabilities Apply Fisher Vector to each individual result to obtain vectors that are the same size Use those fisher vectors for 11 SVM’s (one for each color) COMPLETE STEPS Using Ebay Data (omitting binary images) Use all Google Data (from 30 to 100 images per color) Increase cluster size in GMM from 10 to 128 SVM LINEAR SVM First tried it with libsvm code MATLAB Function: svmtrain (Training, Group) Training: Data to be processed (transpose matrix) Group: Specifies +1 or -1 data Use SVM for each color (black, blue, brown, green, grey, orange, pink, purple, red, white, yellow) Changed to fitcsvm(X,Y) SVM TRAIN OUTPUT FITCSVM OUTPUT CLASSIFY DATA MATLAB Function: svmclassify(SVMStruct,Sample) Use SVMStruct from svmtrain (from each color) Sample: Concatenated Ebay Fisher Vectors Changed to predict(SVMModel, X) SVMModel from fitcsvm SVM CLASSIFY OUTPUT Column vector with the same number of rows as Sample. Each entry (row) in Group represents the class of the corresponding row of Sample. PREDICT OUTPUT Returns Label and Score CURRENT PROGRESS CALCULATE PRECISION Calculate 12 highest scores for each color, using first column only Determine if each score is a correct match by checking indices Calculate each color’s precision FUTURE GOALS Fix Binary function Try process with new data set Data set available: Fahad Shahbaz Khan, Rao Muhammad Anwer, Joost van de Weijer, Andrew D. Bagdanov, Maria Vanrell, Antonio M. Lopez Image retrieval test Object Detection