Herding Generalizes Diverse M -Best Solutions
This work identifies and addresses a flaw in a method for diverse solution extraction in structured prediction, which is incremental but improves reliability for applications like semantic segmentation.
The paper shows that the divMbest algorithm for extracting diverse solutions from Conditional Random Fields is a specific instance of Herding, revealing that it imposes implausible constraints that can lead to incorrect assumptions in certain problem settings. Experiments in semantic segmentation demonstrate that this insight enables the development of better alternatives to address these constraints.
We show that the algorithm to extract diverse M -solutions from a Conditional Random Field (called divMbest [1]) takes exactly the form of a Herding procedure [2], i.e. a deterministic dynamical system that produces a sequence of hypotheses that respect a set of observed moment constraints. This generalization enables us to invoke properties of Herding that show that divMbest enforces implausible constraints which may yield wrong assumptions for some problem settings. Our experiments in semantic segmentation demonstrate that seeing divMbest as an instance of Herding leads to better alternatives for the implausible constraints of divMbest.