Conditional Deep Unrolling for LiDAR Super-Resolution and Improved Segmentation: An End-to-End Solution

Abstract

light detection and ranging (LiDAR) segmentation is essential for autonomous driving, but its accuracy heavily depends on the sensor resolution. While high-resolution LiDAR offers precise environmental detail, its cost is prohibitive. Low-cost sensors, by contrast, produce sparse point clouds that degrade segmentation quality. Recent LiDAR super-resolution (SR) methods aim to address this, but face three key challenges: 1) they are optimized independently of segmentation, leading to suboptimal results; 2) they require costly postprocessing to remove outliers; and 3) they introduce significant computational overhead, often surpassing the complexity of the segmentation network. To address these challenges, we propose a novel conditional optimization problem for LiDAR SR, inspired by recent advances in conditional generative models, which use external guidance signals to influence the generation process. In our case, semantic priors are extracted from the low-resolution data to guide the reconstruction toward semantically meaningful regions. This conditional formulation is efficiently solved using a model-based deep unrolling strategy, yielding a highly efficient conditional SR network. To further enhance performance, the resulting SR model is seamlessly integrated with a high-resolution segmentation network to form a unified, end-to-end trainable architecture. Experiments on public benchmarks show that our method outperforms state-of-the-art LiDAR SR approaches, particularly in preserving semantically important structures. In addition, we successfully deploy the proposed framework on our demo vehicle equipped with an NVIDIA Orin and a 16-channel LiDAR sensor, confirming its practical applicability in real-world driving environments. The code will be available at the project repository.¹¹

http://github.com/alexandrosgk/Conditional-DU-3D-segmentation