Shift R-CNN: Deep Monocular 3D Object Detection with Closed-Form Geometric Constraints
This addresses the problem of accurate 3D object detection from single images for autonomous driving, representing an incremental improvement by integrating geometry into deep learning.
The paper tackled monocular 3D object detection by proposing Shift R-CNN, a hybrid model combining deep learning with geometric constraints, achieving top results on the KITTI benchmark as the best monocular method without pre-trained depth estimation.
We propose Shift R-CNN, a hybrid model for monocular 3D object detection, which combines deep learning with the power of geometry. We adapt a Faster R-CNN network for regressing initial 2D and 3D object properties and combine it with a least squares solution for the inverse 2D to 3D geometric mapping problem, using the camera projection matrix. The closed-form solution of the mathematical system, along with the initial output of the adapted Faster R-CNN are then passed through a final ShiftNet network that refines the result using our newly proposed Volume Displacement Loss. Our novel, geometrically constrained deep learning approach to monocular 3D object detection obtains top results on KITTI 3D Object Detection Benchmark, being the best among all monocular methods that do not use any pre-trained network for depth estimation.