Abstract
Due to hardware limitations, point clouds collected by Li-DAR devices are sparse, making it challenging to locate faraway objects accurately. In this paper, we propose an RGB-guided local point cloud completion network, which aims to improve off-the-shelf 3D object detectors by selectively densifying the collected point clouds. Rather than predicting per-pixel depth in 2D images and projecting them back to pseudo-3D point clouds, our proposed method directly predicts the existence of points in 3D space around input points. Towards this goal, we create a semi-dense labeled local points completion dataset and design a new loss for training the network in a semi-supervised manner. Extensive experiments show that the proposed method can produce reasonable and accurate dense 3D point clouds from sparse inputs, improving off-the-shelf 3D object detectors on the KITTI 3D detection benchmark. The source code of our method will be available at https://github.com/emdata-ailab/LPCC-Net.