Confidence-Guided LiDAR Depth Completion for Robust 3D Object Detection

LiDAR depth completion aims to reconstruct dense depth images from sparse LiDAR scans. While prior methods claim to improve 3D scene understanding, their dense depth maps often introduce erroneous points around object boundaries when directly back-projected into 3D space, which degrades downstream recognition performance. Crucially, these methods focus solely on 2D reconstruction without mechanisms to control pseudo-LiDAR quality. To address this, we propose a novel depth-completion network that jointly predicts a dense depth map and a per-pixel confidence map. The confidence map enables selective back-projection by filtering out unreliable estimates during pseudo-LiDAR point cloud generation. Furthermore, as LiDAR returns exhibit increasing sparsity at longer ranges, our confidence map learns to prioritize reliable long-range estimates. This learned compensation mechanism mitigates sensor limitations while enhancing detection performance for distant objects. On the KITTI 3D benchmark, detectors trained on our pseudo-LiDAR outperform those using raw 32-channel data by 14.2% in 3D AP, narrowing the gap to full 64-channel sensors by 63%.