Incremental Visual Hull Reconstruction

Visual hull reconstruction is the process of generating three dimensional scene geometry from silhouette images captured from different views of a scene. The visual hull is the shape maximally consistent with its silhouette projection [4]. It has been popular due to its good approximating qualities for many objects and its ease and speed of implementation. It is commonly used as a starting point for more elaborate 3D reconstruction methods [5] and finds application in many real-time 3D reconstruction systems due to its good performance [3, 6]. However, current reconstruction systems do not take temporal consistency in the scene into account. In contrast to these approaches, we propose an incremental voxel based visual hull reconstruction method for dynamic scenes using ray casting. By exploiting the fact that there usually occur only small changes between consecutive frames, we can reduce the runtime and calculation complexity in these environments. This allows us to perform real-time reconstruction on a standard processor without having to parallelize the computations. In dynamic scenes such as those captured by multi-camera systems the changes between frames are limited by the speed at which the objects in the scene move or deform. Hence, the number of voxels that change in the reconstruction does not alter dramatically. Therefore it is inefficient to reconstruct each frame independently. Theoretically the only voxels that have to be updated to transform the previous reconstruction into the current one, are the ones with changed occupancy. Here, we explain our algorithm for incremental visual hull reconstruction in dynamic scenes, which uses ray casting to update the visual hull. The first step for updating the reconstruction is to compute the difference images from silhouettes. This allows us to determine which pixels were added or removed from the foreground. Then for each changed pixel in the difference images, we create a ray from the optical center of the camera passing through the pixel and traverse all voxels, which the ray passes, using ray casting [1].