Proportion Priors for Image Sequence Segmentation Claudia Nieuwenhuis, etc. ICCV 2013 Oral
Motivation Current algorithms about Image Sequence Segmentation – Shape similarity Assumption: Rigid body transformation from similar viewpoint Reality: viewpoint changes, articulations or non-rigid deformation – Color similarity Assumption: Similarity of color or feature distributions Reality: Similar or overlapping color distributions between objects and background, Illumination changes – Other methods with relaxed assumptions? Object subspaces or region correspondences, etc. Problem: Optimization problems are complex and hard to solve.
Contribution What property of objects could be preserved among various images in a sequence? Contributions: – Propose framework of proportional preserving priors, add ratio constraint to the classification model. – Construct a convex scheme to approximate it and calculate it efficiently. Invariant and robust to non-rigid deformation, articulation, illumination changes, color overlap Proportional information: Relative size of object parts, eg., size ratio of head to entire body
Problem Definition Bayesian inference for segmentation – : input image of a sequence on the domain – Task of segmentation: Partition the image plane into n pairwise disjoint regions – Compute a labeling Observation likelihood: Color model learned from images Key point
Framework of Proportion Preserving Priors Conditional independence assumption Ratio constraint of one part to whole object: Short boundary length constraint Background, Constant ratio constraint
Proportion Preserving Priors Uniform Distribution Prior Laplace Distribution Prior – Penalize deviations of the ratios from their median Advantage: simple and convex Weak point: not robust to outliers Advantage: perform better and robust to outlier Weak point: not convex How to convert it into convex problem? See paper to get detail.
Results