Wide-baseline multi-view video segmentation for 3D reconstruction

Abstract

Obtaining a foreground silhouette across multiple views is one of the fundamental steps in 3D reconstruction. In this paper we present a novel video segmentation approach, to obtain a foreground silhouette, for scenes captured by a wide-baseline camera rig given a sparse manual interaction in a single view. The algorithm is based on trimap propagation, a framework used in video matting. Bayesian inference coupled with camera calibration information are used to spatio-temporally propagate high confidence trimap labels across the multi-view video to obtain coarse silhouettes which are later refined using a matting algorithm. Recent techniques have been developed for foreground segmentation, based on image matting, in multiple views but they are limited to narrow baseline with low foreground variation. The proposed wide-baseline silhouette propagation is robust to inter-view foreground appearance changes, shadows and similarity in foreground/background appearance. The approach has demonstrated good performance in silhouette estimation for views up to 180° baseline (opposing views). The segmentation technique has been fully integrated in a multi-view reconstruction pipeline. The results obtained demonstrate the suitability of the technique for multi-view reconstruction with wide-baseline camera set-ups and natural background.