Higher-order inference in conditional random fields using submodular functions
Higher-order and dense conditional random fields (CRFs) are expressive graphical models which have been very successful in low-level computer vision applications such as semantic segmentation, and stereo matching. These models are able to capture long-range interactions and higher-order image statis...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2023
|
| Summary: | Higher-order and dense conditional random fields (CRFs) are expressive graphical
models which have been very successful in low-level computer vision applications
such as semantic segmentation, and stereo matching. These models are able to
capture long-range interactions and higher-order image statistics much better
than pairwise CRFs. This expressive power comes at a price though - inference
problems in these models are computationally very demanding. This is a
particular challenge in computer vision, where fast inference is important and
the problem involves millions of pixels.
In this thesis, we look at how submodular functions can help us designing
efficient inference methods for higher-order and dense CRFs. Submodular
functions are special discrete functions that have important properties from
an optimisation perspective, and are closely related to convex functions. We
use submodularity in a two-fold manner: (a) to design efficient MAP inference
algorithm for a robust higher-order model that generalises the widely-used
truncated convex models, and (b) to glean insights into a recently proposed
variational inference algorithm which give us a principled approach for applying
it efficiently to higher-order and dense CRFs. |
|---|