CustomDepth: Customizing point-wise depth categories for depth completion

Classification-based depth completion methods have achieved remarkable performance. However, the result is still coarse due to the limitation of using unified depth categories to represent depth distribution. In this work, we propose CustomDepth which can customize exclusive depth categories for each image point to boost performance. To this end, CustomDepth introduces a depth subdivision module that allocates adaptive depth categories for each point based on its properties, instead of refining a set of unified categories for all points. With these adaptive depth categories, CustomDepth utilizes a binary classifier to determine whether a point is located in front of or behind each depth category. The classification results are then accumulated using a rendering approach to calculate the final depth result. To reduce computational burden, CustomDepth also incorporates an image subdivision module that selectively processes a subset of error-prone points. Extensive experiments demonstrate that CustomDepth is a lightweight and flexible framework that achieves competitive performance compared to existing classification-based methods.