Past Papers Organized by Projects


Statistical Separability Constraints

In research, the focus is often on discovering absolute constraints. However, for many practical problems, such constraints do not exist. One solution is to rely on the law of large numbers to agglomerate many weak constraints into robust constraints. We show how this approach can provide uniquely robust solutions to difficult problems.

"Dimensionality's Blessing: Clustering Images by Underlying Distribution," CVPR 2018. Statistical analysis of high dimension space permits clustering of images by their underlying distribution;

"GMS: Grid-based Motion Statistics for Fast, Ultra-robust Feature Correspondence," CVPR 2017. Rapid counting algorithm allows match numbers in a region to disambiguate true and false correspondence;

"RepMatch: Robust Feature Matching and Pose for Reconstructing Modern Cities," ECCV 2016. Large number of matches enable disambiguation of repeated structures.


Structure-from-Motion

Structure-from-Motion is the estimation of 3D depth and camera position from two or more views of the same scene. Depth and camera position are typically computed from feature correspondence between image pairs. A perennial problem is that feature correspondence may be wrong or inadequate in number. We show how introducing coherence constraints allows for large numbers of correct feature matches. This provides pose estimation with unprecedented stability and accuracy.

"CODE: Coherence Based Decision Boundaries for Feature Matching," PAMI 2018. Ultra-robust wide-baseline feature correspondence;

"GMS: Grid-based Motion Statistics for Fast, Ultra-robust Feature Correspondence," CVPR 2017. Real-time extension of CODE matching with a statistical separability function.

"RepMatch: Robust Feature Matching and Pose for Reconstructing Modern Cities," ECCV 2016. Extension of CODE matching to Structure-from-Motion with enhanced pose estimation stability;


Feature Matching

Feature correspondence becomes challenging when the view-points change too much. We show the problem can be understood as differentiating true and false matches. Prior techniques use strict similarity thresholds that discard many true matches. However, false matches are usually randomly scattered while true matches tend to be coherent (clustered around a few dominant motions), this creates a statistical separability constraint. The model can filter the entire set of nearest neighbour matches (which typically contains over 90% false matches) for true matches. We integrate our technique into a full feature correspondence system which remain effective where previous techniques provide few or no matches.

"CODE: Coherence Based Decision Boundaries for Feature Matching," PAMI 2018. Ultra-robust wide-baseline feature correspondence, an extension of bilateral functions;

"GMS: Grid-based Motion Statistics for Fast, Ultra-robust Feature Correspondence," CVPR 2017. Real-time extension of CODE matching with a statistical separability function;

"Bilateral Functions for Global Motion Modelling," ECCV 2014 The first feature matching formulation to reliably separate true and false matches in extreme conditions;

"Robust Non-parametric Data Fitting for Correspondence Modelling," ICCV 2013. Convex formulation for finding the smoothest, non-parametric curve through data-points. Foundational work for the later papers.