Ghunhui Gu, Joseph J. Lim, Pablo Arbeláez, Jitendra Malik

University of California at BerkeleyBerkeley, CA 94720

IntroductionApproachExperimental ResultsConclusion

Introduction

Early work in the late 90s , the domain strategy for object detection in a scene has been multi-scale scanning

: is there an instance of object category C in the window?

It differs significantly from the nature of human visual detection

So,

This paper focus on using regions , which have some properties:

(1)They encode shape and scale information of objects naturally

(2)They specify the domains on which to compute various features, without being affected by clutter from outside the region (background)

(3)But its not popular as features due to their sensitivity to segmentation error

ApproachOverview the method

Framwork for Region weighting

Main recognition algorithm(1)Voting(2)Verification(3)Segmentation

The “bag of regions” representation of a mug example

[2] P. Arbel´aez, M. Maire, C. Fowlkes, and J. Malik. From contoursto regions: An empirical evaluation. In CVPR, 2009.

All node generated by[2]

Region cues:

Contour shape, given by the histogram of oriented responses of the contour detector gPb [22]

Edge shape, where orientation is given by local image gradient (by convolution)

Color, represented by the L*, a and b histograms in the CIELAB color space

http://en.wikipedia.org/wiki/Lab_color_space

Texture, described by texton histograms

Describe a region by subdividing evenly its bounding box int an n x n grid

(a)Original image, (b) A region from the image, (c) gPb [22]Representation of the region in (b), (d) Our contour shape descriptor based on (c)

The “contour shape” region descriptor

[22] M. Maire, P. Arbel´aez, C. Fowlkes, and M. Malik. Usingcontours to detect and localize junctions in natural images.In CVPR, 2008.

Discriminative Weight Learning

I and J are objects of same category, but K is an object of a different category

Discriminative Weight Learning

The pipeline of object recognition algorithm

Voting , Verification, Segmentation three stage

Voting stage

This transformation provides not only position but also scale estimation of the object. It also allows for aspect ratio deformation of bounding boxes.

Voting

Vote of bounding box of the object(Transformation function )

Vote score

Transformation function model they use

Given a query image and an object category, is to generate hypotheses of bounding boxes and support of objects of that category in the image

Verification

The verification score

The average of the probabilities

The overall detection score --Product of the two score

Segmentation

Green for object and Red for background

To recover the complete object support from one of its parts

Experimental Results

1. ETHZ shape2. Caltech-101

Data base:

Detection performance

ETHZ shape

Region tree : on average ~ 100 regions per image

Color and texture are not very useful in this data base

Choose the functions in Eqn.11 as:

Split the entire set in to half training and half test for each category

ETHZ shapes

Caltech 101Randomly pick 5, 15 or 30 images for training and up to 15 images in disjoint set for test

Geometric blur[4]

Caltech 101

conclusionPresented a unified framework for object

detection, segmentation, and classification using regions.

(1)Cue combination significantly boosts recognition performance

(2)Reduces the number of candidate bounding box by order of magnitude over standard sliding window scheme due to robust estimation of object scales from region matching