10.1.1.61.7150

10.1.1.61.7150 - The Layout Consistent Random Field for...

Info iconThis preview shows pages 1–2. Sign up to view the full content.

View Full Document Right Arrow Icon
The Layout Consistent Random Field for Recognizing and Segmenting Partially Occluded Objects John Winn Microsoft Research Cambridge Cambridge, UK jwinn@microsoft.com Jamie Shotton Department of Engineering University of Cambridge, UK jdjs2@cam.ac.uk Abstract This paper addresses the problem of detecting and seg- menting partially occluded objects of a known category. We Frst deFne a part labelling which densely covers the object. Our Layout Consistent Random ±ield (LayoutCR±) model then imposes asymmetric local spatial constraints on these labels to ensure the consistent layout of parts whilst allow- ing for object deformation. Arbitrary occlusions of the ob- ject are handled by avoiding the assumption that the whole object is visible. The resulting system is both efFcient to train and to apply to novel images, due to a novel annealed layout-consistent expansion move algorithm paired with a randomised decision tree classiFer. We apply our technique to images of cars and faces and demonstrate state-of-the-art detection and segmentation performance even in the pres- ence of partial occlusion. 1. Introduction This paper addresses the problem of detecting and seg- menting both clean and partially occluded deformable ob- jects of a known category. The approach uses a part la- belling which densely covers the object and models the label distribution using an enhanced Conditional Random Field which we call the Layout Consistent Random Field (LayoutCRF). The use of parts has several advantages. First, recog- nising parts of an object allows for object detection under partial occlusion. Second, there are local spatial interac- tions between parts that can help with detection; for exam- ple, we expect to ±nd the nose just above the mouth on a face. Hence, we can exploit local part interactions to ex- clude invalid hypotheses at a local level. Third, knowing the location of one part highly constrains the locations of other more distant parts. For example, knowing the loca- tions of wheels of a car constrains where the rest of the car can be detected. Thus, we can improve object detection by incorporating long range spatial constraints on the parts. Fi- nally, by inferring a part labelling for the training data, we can accurately assess the variability in the appearance of each part, giving better part detection and so better object detection. A key aspect of the model is the use of asymmetric pair- wise potentials to capture the spatial ordering of parts, e.g. car wheels must be below car body, not vice-versa. These asymmetric potentials allow propagation of long-range spa- tial dependencies using only local interactions. The pairwise potentials are carefully constructed to dis- tinguish between various types of occlusion, such as object occluding background 1 , background occluding object, and object occluding object. The model is capable of represent- ing multiple object instances which inter-occlude, and in- fers a pairwise depth ordering.
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Image of page 2
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 03/27/2010 for the course CS 123 taught by Professor Darghooz during the Spring '10 term at Albion College.

Page1 / 8

10.1.1.61.7150 - The Layout Consistent Random Field for...

This preview shows document pages 1 - 2. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online