Visual Hull Construction in the Presence of Partial Occlusion

In this paper, we propose a visual hull algorithm, which guarantees a correct construction even in the presence of partial occlusion, while "correct" here means that the real shape is located inside the visual hull. The algorithm is based on a new idea of the "extended silhouette", which requires the silhouette from background subtraction and the "occlusion mask" of the same view. In order to prepare the occlusion mask, we also propose a novel concept of "effective boundary" of moving foreground objects in a video obtained from a static camera. The accumulation of the effective boundary through time automatically gives robust occluder boundaries. We theoretically prove that our algorithm deterministically computes the tightest, correct visual hull in the presence of occlusion. Both synthetic and real examples are given as a demonstration of the correctness of the algorithm. Finally we analyze that this new algorithm is still within the time complexity of the traditional method.

1. Introduction

A visual hull is defined as the intersection of silhouette cones from 2D camera views, which captures all geometric information given by the image silhouettes [1]. One basic property is that a visual hull is the largest volume to be consistent with silhouette information from all views [2]. The actual shape is always contained inside the constructed visual hull. This is referred to as the visual hull conservation constraint. Hence multiple view shape-from-silhouette algorithms for computing the visual hull are widely used to provide a coarse 3D shape estimate which can serve as an initialization for more sophisticated 3D shape reconstruction algorithms, such as [3].