Search papers, labs, and topics across Lattice.
The paper introduces Structured SIR, a novel Sampled Importance Resampling (SIR) algorithm for probabilistic image registration that addresses limitations of variational inference by enabling expressive, multi-modal characterization of uncertainty. The method uses a memory-efficient high-dimensional covariance parameterization as the sum of a low-rank covariance and a sparse, spatially structured Cholesky precision factor. Experiments on 3D brain MRI registration demonstrate that Structured SIR produces better calibrated uncertainty estimates and captures multi-modal posterior distributions compared to variational methods, while maintaining equivalent or better accuracy.
Ditch the overconfident posteriors: Structured SIR offers a memory-efficient way to capture complex, multi-modal uncertainty in high-dimensional image registration, outperforming variational inference.
Image registration is an ill-posed dense vision task, where multiple solutions achieve similar loss values, motivating probabilistic inference. Variational inference has previously been employed to capture these distributions, however restrictive assumptions about the posterior form can lead to poor characterisation, overconfidence and low-quality samples. More flexible posteriors are typically bottlenecked by the complexity of high-dimensional covariance matrices required for dense 3D image registration. In this work, we present a memory and computationally efficient inference method, Structured SIR, that enables expressive, multi-modal, characterisation of uncertainty with high quality samples. We propose the use of a Sampled Importance Resampling (SIR) algorithm with a novel memory-efficient high-dimensional covariance parameterisation as the sum of a low-rank covariance and a sparse, spatially structured Cholesky precision factor. This structure enables capturing complex spatial correlations while remaining computationally tractable. We evaluate the efficacy of this approach in 3D dense image registration of brain MRI data, which is a very high-dimensional problem. We demonstrate that our proposed methods produces uncertainty estimates that are significantly better calibrated than those produced by variational methods, achieving equivalent or better accuracy. Crucially, we show that the model yields highly structured multi-modal posterior distributions, enable effective and efficient uncertainty quantification.
