時空間情報伝搬に基づく多眼動画像の対話的セグメンテーションシステムの構築

概要

論文の詳細を見る
In this paper, we address a problem faced by interactive segmentation, which is the extraction of a specific object region, when applied to multi-view video images. Many effective techniques for interactive segmentation have recently been developed to deal with static images. However, directly applying such methods to multi-view video images is impractical because of the huge number of images. We propose a system in which only a small number of images is processed with a user interface while the many remaining images are processed automatically. In the automatic process, segmentation is performed by first propagating the shape and color information of the object from already segmented images to the currently targeted image, and then minimizing the energy function defined by the propagated information. In our experiment, we captured 200 frames from 25 multi-view cameras (5000 images total), and manually segmented only 5 images. The automatic segmentation of the remaining images was performed successfully.