- Date and time: Tuesday Nov 6, morning session
- Venue: Daejeon Convention Center
- Duration: 4 hours
- Presenters: Tat-Jun Chin (The University of Adelaide), Hanzi Wang (Xiamen University) and Yong Duek Seo (Sogang University). See biographies.
Course abstract
Fitting models onto noisy measurements is a fundamental task in computer vision. The need for robust statistical fitting arises since outliers frequently affect the data. As an important mid-level processing step, accurate and computationally efficient robust statistical fitting is essential to the success of many computer vision applications.
This tutorial is aimed at introducing robust statistics in computer vision to new researchers. We focus on robust fitting on data with severe outlier contamination, data with unknown noise scales, and data with multiple structures. Special attention is devoted to the optimisation of robust criteria, outlier rejection schemes (e.g., under the L0 and L-infinity norm), and model selection. The topics covered will be put in the context of applications such as projective estimation and motion analysis.
Course topics
- Section 1 - slides in pdf (6.1MB)
-
- Course organisation and plan
- Problem statement and applications in computer vision
- Basic theory and methods
- Robust criteria and breakdown points
- Random sampling and related heuristics
- Random sampling under multiple structures
- Section 2 - slides in pdf (566KB)
-
- M-estimators and iteratively reweighted least squares
- Least absolute deviation and least maximum deviation
- Least median and least K-th order
- Section 3 - slides in pdf (33.8MB)
-
- Outlier removal based on L-infinity norm minimisation
- Application to 3D reconstruction
- L0 norm minimisation for consensus maximisation
- Application to shape-from-shading
- Rotation search algorithms for nonlinear models
- Section 4 - slides in pdf (2.5MB)
-
- Scale estimation and segmentation
- Multi-structure fitting and model selection
- Hough Transform and mean shift
- Spectral clustering and other recent methods
- Higher order constraints for multi-structure fitting
- Energy minimisation with label costs
Affiliations:
