PASCAL Visual Object Classes Challenge 2006

People:

Dhruv Batra, Gunhee Kim

Keywords:

CRFs, DRFs, MRSC, Loopy BP

Class Project:

16-721 Advanced Robot Perception

Instructor:

Alyosha Efros

Abstract

We draw a distinction between two kinds of classes in the VOC2006 database, first kind being the “structured” classes (car, bus, bicycle, motorbike), the second being “unstructured/deformable” classes (person, cat, cow, horse, dog, sheep). We make an observation that while the former possess strict geometry which is rarely deformed, the latter can be treated as a texture recognition problem with consistency in their background acting as context. We work on two different algorithms to harness these two consistencies.

Method 1 (for “structured” classes). We pose this problem as a local feature matching problem between the test images and the annotated training images (Hand segmented to exact boundaries by us). The feature detector used was Lowe’s DOG [1]. The local features we experimented with were Yan Ke’s PCA-SIFT [2] and actual patches cut out of images at proper scales in Gaussian pyramids. To boost the matching over a NN based scheme, and to incorporate spatial and geometric constraints in the matched local features, we used Leordeanu and Hebert's spectral correspondence [3] based matching. Scores were generated which denoted consistency of matched features. This method, although robust to small intra-class shape variances, requires a certain class geometry to be preserved, and this is precisely why we cannot use this method for highly deformable classes.

Report:

Dhruv Batra, Gunhee Kim, Alyosha Efros (Team: AP06_Batra). The Pascal Visual Object Classes Challenge 2006 (VOC2006) Results.
[ pdf | challenge homepage | online results ]

Figure 1: Example image from PASCAL VOC 2006