Silhouettes Fusion for 3D Shapes Modeling with Ghost Object Removal

In this paper, we investigate a practical framework to compute a 3D shape estimation of multiple objects in real-time from silhouette probability maps in multi-view environments. A popular method called Shape From Silhouette (SFS), computes a 3D shape estimation from binary silhouette masks. This method has several limitations: The acquisition space is limited to the intersection of the camera viewing frusta; SFS methods reconstruct some ghost objects which do not contain real objects, especially when there are multiple real objects in the scene; Lastly, the results depend heavily on quality of silhouette extraction. In this paper we propose two major contributions to overcome these limitations. First, using a simple statistical approach, our system reconstructs objects with no constraints on camera placement and their visibility. This approach computes a fusion between all captured images. It compensates for bad silhouette extraction and achieves robust volume reconstruction. Second, a new theoretical approach identifies and removes ghost objects. The reconstructed shapes are more accurate than current silhouette-based approaches. Reconstructed parts are guaranteed to contain real objects. Finally, we present a real-time system that captures multiple and complex objects moving through many camera frusta to demonstrate the application and robustness of our method.