Are Spatial and Global Constraints Really Necessary for Segmentation?
Many state-of-the-art segmentation algorithms rely on Markov and
Conditional Random Fields designed to enforce spatial and global
consistency constraints. This often results in fairly complex
designs. As a result, estimating the parameters or computing the best
Maximum-A-Posteriori (MAP) assignment for such models become a
computationnaly expensive task.
In this paper, we argue that similar levels of performance can be achieved on
the PASCAL and MSRC datasets using a much simpler design that essentially
ignores those constraints. It replaces them by global features that leverage
evidence from the whole image and uses them to bias the preference of individual
pixels.
This does not prove that spatial and consistency constraints are not important
but points to the conclusion that they should be validated in a larger context.
