Provide a summary for "RGB-(D) Scene Labeling: Features and Algorithms" paper, written by X Ren, L Bo, D Fox - Computer Vision and Pattern Recognition 2012 - ieeexplore.ieee.org
2. Introduction
Scene labeling challenges
Pipeline
Feature Extraction
Super-pixel formulation and classification
Classifying segmentation tree paths
Classifying super-pixels MRF
Datasets and results
Agenda
3. Scene Labeling
Labeling of each pixel in an image to a certain class
Scene Labeling can be done
Indoors
Label a Sofa in a Bedroom
Label a door in a living room
Outdoors
Label a car in street
Label building in street
Scene Labeling
6. Indoor scene labeling challenges
Large variations of scene types
Lack of distinctive features
Poor illumination
Scene Labeling
7. Benefits of using depth feature in scene labeling
Increased accuracy and robustness
Body pose estimation
3D mapping
Object recognition
3D modeling and interaction
Scene Labeling
9. 1. Extract features using Kernel descriptor (KDES).
2. Aggregate descriptors in dense region into super-
pixels using Efficient match kernels (EMK)
3. Classify super-pixels using Linear support vector
machine (SVM)
4. Label super-pixels by classifying paths of
segmentation tree.
5. Label super-pixels using super-pixel MRF
Pipeline
10. Kernel Descriptors (KDES), a unified framework that
uses different aspects of similarity (kernel) to derive
patch descriptors.
Image gradient
Spin/normal
Color
Depth gradient
Features Extraction (Step 1)
11. Efficient match kernels (EMK) to transform and
aggregate descriptors in a set S (grid locations in the
interior of a superpixel ‘s’).
Super-pixels are not of the same size.
Super-pixel formation (Step 2)
12. Linear Support vector machine (SVM)
Non-probabilistic binary linear classifier.
Classify superpixels (Step 3)
16. Classifying paths in segmentation tree
If we accumulate features over paths, the accuracy
continues to increase to the top level
The initial part of the curves overlap, suggesting there is
little benefit going to superpixels at too fine scales
Contextual Models
18. Superpixel MRF with gPb
standard MRF formulation. We use Graph Cut to find
the labeling that mini- mizes the energy of a pairwise
MRF
Contextual Models (Step 5)
22. Rgb-(d) scene labeling: Features and algorithms
X Ren, L Bo, D Fox - Computer Vision and Pattern
Recognition 2012 - ieeexplore.ieee.org
Context by region ancestry
JJ Lim, P Arbeláez, C Gu, J Malik - Computer Vision, 2009
IEEE 2009 - ieeexplore.ieee.org
References