Leveraging Proposals and Shape Reconstruction for Monocular 3D Object Detection and Object-centric Stereo Matching for 3D Object Detection

Safe autonomous driving requires reliable 3D object detection-the task of estimating 3D bounding boxes for objects of interest in the scene. This thesis introduces a monocular 3D object detector that accurately estimates the 3D position of objects by first generating 3D proposals, using a novel 2D bounding box prior, then refining them in a second stage. Depth information is further learned by jointly optimizing these predicted object positions and the task of point cloud reconstruction. A stereo 3D object detector is also introduced. Compared to previous methods, it only estimates disparities for pixels belonging to objects of interest, which mitigates point cloud artifacts seen in full scene depth maps and improves runtime speed. Moreover, a novel point cloud loss is presented to combat the issue that typical disparity loss functions insufficiently penalize far depths. Both detectors achieve state-of-the-art results on the KITTI dataset when compared to other camera-based methods.M.A.S