Fusion of depth maps based on confidence

In this paper, we propose a confidence-based depth maps fusion approach for multi-view 3D reconstruction. We present a four-stage method to refine the fused depth map to get a higher accuracy. At first, we utilize the off-the-shelf Kinect device to acquire the raw multi-view depth maps quickly. To improve the accuracy, we compute the confidence values for the estimated depth maps in the second stage to find a theoretically optimal value. Next, we discard some outliers within the confidence region to further optimize the depth maps. Finally, we attempt to fill the holes to refine the fused depth map. Similarly, we will point out some limitations of our confidence-based algorithm in obtaining the accurate depth map when putting into practice.