CU-Net: LiDAR Depth-Only Completion With Coupled U-Net

LiDAR depth-only completion is a challenging task to estimate dense depth maps only from sparse measurement points obtained by LiDAR. Even though the methods have been widely developed, there still significant performance gap with RGB-guided that utilize extra color images. We find existing can obtain satisfactory results in areas where are almost accurate and evenly distributed (denoted as normal areas), while limited foreground background overlapped due occlusion overlap areas) no around blank since reliable input information these areas. Building upon observations, we propose an effective Coupled U-Net (CU-Net) architecture for completion. Instead of directly using large network regression, employ local values provide global initial The predicted two coupled U-Nets fused learned confidence final results. In addition, confidence-based outlier removal module, which removes outliers simple judgment conditions. Our proposed method boosts fewer parameters achieves state-of-the-art on KITTI benchmark. Moreover, it owns powerful generalization ability under various densities, varying lighting, weather