PROJECT TITLE :

CenterNet3D An Anchor Free Object Detector for Point Cloud

ABSTRACT:

In autonomous driving, one of the most important tasks is the quick and accurate detection of three-dimensional objects from point clouds. The currently available one-stage 3D object detection methods are capable of achieving real-time performance; however, they are dominated by anchor-based detectors, which are inefficient and require additional post-processing. In this article, we get rid of anchors and model an object as a single point, which is the point in the middle of its bounding box. We propose an anchor-free network called CenterNet3D that can perform 3D object detection without using any anchors. This network is based on the center point. Our CenterNet3D algorithm directly regresses 3D bounding boxes and makes use of keypoint estimation in order to locate center points. However, due to the inherent sparsity of point clouds, the center points of 3D objects are likely to be in empty space, which makes it difficult to estimate the boundaries of the object accurately. We propose adding an additional corner attention module to the CNN backbone so that it will be forced to pay more attention to object boundaries so that we can resolve this issue. In addition, taking into consideration that one-stage detectors are susceptible to the problem of discordance between the predicted bounding boxes and the corresponding classification confidences, we devised an effective keypoint-sensitive warping operation with the goal of aligning the confidences with the predicted bounding boxes. Because there is no non-maximum suppression in our recently proposed CenterNet3D, it is both more effective and easier to use. We evaluate CenterNet3D using the KITTI dataset, which is quite popular, as well as the nuScenes dataset, which is more difficult. Our method achieves better results than any other state-of-the-art anchor-based one-stage method and achieves results that are comparable to those of two-stage methods as well. It achieves the best speed and accuracy trade-off possible and has an inference speed of 20 frames per second (FPS).