Browse Prior Art Database

Region of Interest Optimization for Improved Object Classification

IP.com Disclosure Number: IPCOM000182029D
Original Publication Date: 2009-Apr-22
Included in the Prior Art Database: 2009-Apr-22
Document File: 1 page(s) / 19K

Publishing Venue

IBM

Abstract

The present invention is a novel method for visual object classification and verification in retail stores. Our method relies on an algorithm which determines a discriminative region of interest (based on an optimization process) for an object and performs classification based only on the selected region.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 90% of the total text.

Page 1 of 1

Region of Interest Optimization for Improved Object Classification

We are interested in solving the problem of visual object classification/verification to prevent fraud in retail stores. In particular, ticket switching is a common type of fraud in which thefts steal expensive items by replacing their barcodes with barcodes of cheap products. The classification problem in this scenario is very challenging due to the high number of products and uncontrolled variations in lighting and pose of the objects. Object segmentation problems are another major source of error for the classification task. In many cases, the segmentation result includes parts of the image belonging to the background, posing serious problems to model the appearance of the objects.

Our invention relies on the observation that often only a specific part of the object is sufficent to discriminate it from other objects. Given a noisy object segmentation mask, our goal is to find a region of interest (ROI) inside the object that is optimal for classification. Therefore, even if the segmentation includes parts of the background, after the ROI optimization, we expect these parts to be discarded.

The ROI optimization occurs during training and testing time. During training, a ROI is optimized for each exemplar to minimize the classification loss, taking into account other constraints (like the size of the region and the confidence of the segmentation). During t...