This preview has intentionally blurred sections. Sign up to view the full version.View Full Document
Unformatted text preview: HW4 Report Zhongzang Lin Mar 15, 2010 1 Introduction This homework is about object recognition using the bag-of-words model, testing on the Graz-02 images of bikes, cars, and people. Images of each classs are divided into training set and test set. Use the training sets to learn the words in the bag as well as the representation of the object, and then use them to classify test images. The basic idea is first to divide each training image into patches either with regular grid or in a hierarchi- cal way, compute the Histogram of Gradient (HOG) features for the patches, find the codewords with K-mean; then compute the occurence of the codewords in each class, the histograms are the representation of the classes. At last, compute the histogram and classify each test image by finding the nearest neighbor with χ 2 distances. 2 Literature Review There are various ways in object recognition and image classification. Brown et al.  use techniques based on invariant local features to select matching images, and a prob- abilistic model for verification, which is used in automatic construction of panoramas. Fergus et al  use a probabilistic representation for all aspects of the object: shape, appearance, occlusion and relative scale. They use entropy-based feature detector to select regions and a Bayesian model to classify images. Webe  treats objects as flexible constellations of rigid parts and represent the variability as a joint probability density function and learn the constellation models in a unsupervised fashion. 3 Description of Approach The whole classification process is divided into two phases: learning and classification. The first phase is dedicated to find the codewords, and compute the description of each class; in the second phase, features for test images are calculated and compared to description of the classes, and the image is classified as the nearest class....
View Full Document
- Spring '08
- Machine Learning, Type I and type II errors, localization error, training images