沿法线方向传播的深度图像补洞算法

密集的深度信息在计算机视觉的各种任务中有广泛的应用，然而深度相机在有光泽、透明、较远处的物体表面通常无法探测到深度信息，映射到深度图片上形成了大小不一的孔洞。因此，提出了一种沿法线方向传播的单一深度图像补洞算法。本文方法沿着物体本身的曲面变化进行扩散，把深度图片的补全问题转化成几何完成的问题。首先把2D的深度图像扩展到3D点云，然后3D点云沿着孔洞边界的法线方向向内收缩。收缩的过程中加入类似正态滤波器的约束函数来模拟深度的变化，使填充的点云更加适合整个物体的结构。最后把3D点云重新映射到2D图片上。本文的算法在NYU-Depth-v2数据集上进行测试，实验证明本文算法对孔洞的填充有较好的效果。

Depth Image Hole Filling Algorithm Propagating Along the Normal Direction

Dense depth information is widely used in various tasks in computer vision. However, depth cameras usually cannot detect depth information on glossy, transparent, and distant object surfaces, so holes of various sizes are formed on the depth image. Therefore, a depth image hole filling algorithm propagating along the normal direction is proposed in this paper. The method in this paper spreads along the surface changes of the object itself, and transforms the problem of depth completion into the problem of geometric completion. Firstly, the 2D depth image is expanded to a 3D point cloud, and then the 3D point cloud is contracted inward along the normal direction of the hole boundary. In the shrinking process, a constraint function similar to a normal filter is added to simulate the change of depth, so that the filled point cloud is more suitable for the structure of the whole object. Finally, the 3D point cloud is remapped to the 2D image. The algorithm in this paper is tested on the NYU-Depth-v2 dataset, and the experiments show that the algorithm has a good effect on hole filling.