Abstract

We propose a graph cut model to optimize bidimensional shapes with respect to the elastica energy. At each iteration our model selects the shape of minimum elastica value among a set of candidates generated by a discrete process that we call the balance coefficient flow. In this work we show how the balance coefficient flow is related to the curve-shortening flow and how our model can be used to execute the task of image segmentation. Finally, we compare our segmentation model with the classical graph cut algorithm.

Authors: Daniel Martins Antunes, Jacques-Olivier Lachaud and Hugues Talbot.

Report: HAL pre-print

GitHub Project

The gf-flow project implements the graph-flow algorithm and provides a console and GUI interface to interact with. The code is written in C++ and it is easily setup via a docker container.

Coco-Dataset Experiment Report

The graph-flow segmentation model is compared against the grabcut algorithm in a subset of the coco-dataset. The results for all the images in the experiment are available in this report.

Paper Figures

All the figures in the paper can be regenerated via scripts contained in the gf-paper-figures project