NegCut: Automatic Image Segmentation based on MRF-MAP
This work addresses the problem of enabling automatic image segmentation for computer vision applications, though it appears incremental as it builds on existing MRF-MAP methods.
The paper tackles the challenge of automating MRF-MAP image segmentation, which typically requires user interaction, by proposing NegCut, an approximation algorithm that uses minimum cuts on graphs with negative weights and achieves competitive segmentation quality in experiments.
Solving the Maximum a Posteriori on Markov Random Field, MRF-MAP, is a prevailing method in recent interactive image segmentation tools. Although mathematically explicit in its computational targets, and impressive for the segmentation quality, MRF-MAP is hard to accomplish without the interactive information from users. So it is rarely adopted in the automatic style up to today. In this paper, we present an automatic image segmentation algorithm, NegCut, based on the approximation to MRF-MAP. First we prove MRF-MAP is NP-hard when the probabilistic models are unknown, and then present an approximation function in the form of minimum cuts on graphs with negative weights. Finally, the binary segmentation is taken from the largest eigenvector of the target matrix, with a tuned version of the Lanczos eigensolver. It is shown competitive at the segmentation quality in our experiments.