Depth Super-Resolution for 3D Object Detection in Mobile Robotics

The problem of 3D object detection is to detect objects, presented in the scene, and estimate their boundaries and orientations. Range information of the scene is essential together with visual data for accurate scene analysis. Nowadays such data is simple to collect from the environment with help of various scanners. However, in most of the cases the resolution of this scans are very limited what leads to a poor performance of data processing methods. The common sense suggests that usage of super-resolved data, that is the data with computationally increased level of detail, should improve the performance of a 3D object detection algorithms. It was demonstrated that task-driven photo super-resolution improves accuracy of 2D object detection. At the same time, to the best of our knowledge, there was no corresponding evidence for range-data. This work is the first step in the investigation of suggestion, and it aims to become a foundation for future research. Here implementations of modern methods of super-resolution and 3D object detection and experiments with them are presented.