Designing effective samplers
In this week we will look at how we can design effective samplers for fitting algorithms. We start by discussing typical problems that occur when fitting models, how to detect them, as well as simple mitigation strategies based on parameter tuning, block-wise proposals and mixture proposals. We then discuss the Metropolis-Filter proposal, which provides an effective way to filter complex proposal based on simpler parts. Finally we discuss how we formulate the Iterative Closest Point and Active Shape Modelling algorithm as a proposal. This allows us to use these classical, deterministic algorithms as a part in our more complex, fully probabilistic setting.
Steps
- Detecting problems in the sampler (Article)
- Simple sampling strategies for shape model fitting (Article)
- Designing effective chains using the Metropolis-Filter proposal (Article)
- Directed proposals using ICP and ASM steps (Article)
- A closest point proposal for MCMC-based Probabilistic Surface Registration Video (ECCV 2020)
Questions and remarks
Please use our Gitter Chat to ask questions, discuss with your peers or to give us feedback.