Mentor : Prof. Amitabha Mukerjee Learning to Detect Salient Objects Team Members - Avinash Koyya Diwakar Chauhan
Salient Object : Object of visual attention Foreground object Familiar but unknown Aim : To detect such object What is it?
Fitting image to smaller screen Image Collection browsing Tracking objects Why is it?
Itti’s algorithm Colors, Intensity, orientation Require parameter setting for good result Matlab saliency toolbox “Learning to Detect Salient object” by Tie Liu, 2011 Previous work
Binary classification of each pixel Bottom up model Approach Salient Object Features Pairwise Feature Conditional Distribution Linear Combination CRF Learning Salient Background
Multiscale Contrast Center -Surround Histogram Color spatial Distribution Salient object features
Sudden sharp change of intensity? Simulation of human visual receptive field Can carry the point of attention Multiscale Contrast ImageGaussian Filter Contrast map Rescale to ½ of previous Filtered Image Multiscale Contrast Map and sumRescale
Examples Input Output Intermediate Contrast outputs in Pyramid
Regional feature. Attention proportional to distinction from the surroundings. Center Surround Histogram Integral Histogram Chi-Square distance Feature map Aspect ratios ; Sizes Maximize Weighted Sum
Center Surround histogram distances with different locations and sizes.
Original Image Output claimed by the paper Output we obtained
Global feature. Wider a colour is distributed in the image, the less possible it is that a salient object contains this colour. Colour Spatial Distribution K-means Gaussian Mixture Models Spatial Variance EM algorithm Feature MapCenter Weight Weighted Sum
Original Image Output claimed by the paper Output we obtained
We are using MSRA salient object database 20,000 training images and 5,000 test data Training images are labeled by three people Database
Process of learning from the images Learn the linear weights in under maximum likelihood criteria CRF Learning
Feature maps ready Labeling of database and CRF learning in progress So far…