A Comprehensive Survey on Deep Learning-Based LiDAR Super-Resolution for Autonomous Driving

2026-02-19 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

This paper presents the first comprehensive survey of deep learning-based LiDAR super-resolution (SR) methods for autonomous driving, addressing the challenge of sparse point clouds from affordable low-resolution sensors. It categorizes existing approaches into four main types: CNN-based architectures, model-based deep unrolling, implicit representation methods, and Transformer and Mamba-based approaches. The survey establishes fundamental concepts, including data representations, problem formulation, benchmark datasets like KITTI and nuScenes, and evaluation metrics such as MAE, Chamfer Distance, and IoU. Current trends emphasize range image representation, extreme model compression, resolution-flexible architectures, real-time inference (targeting over 25 fps), and cross-sensor generalization. The authors highlight that while significant progress has been made, challenges like cross-sensor generalization and ensuring downstream task performance persist.

Key takeaway

For AI Scientists and Computer Vision Engineers developing autonomous driving systems, understanding the trade-offs between LiDAR SR architectures is crucial. Prioritize Transformer and Mamba-based methods for superior geometric consistency and real-time performance, but be mindful of the persistent challenge of cross-sensor generalization. Focus future development on hybrid domain processing, self-supervised learning, and multi-modal fusion to create more robust and sensor-agnostic SR solutions, ultimately reducing sensor costs while maintaining safety standards.

Key insights

LiDAR super-resolution uses deep learning to enhance sparse point clouds, enabling cheaper sensors to perform like high-resolution ones.

Principles

• Range image representation is efficient for LiDAR SR.
• Model-based deep unrolling improves parameter efficiency.
• Implicit functions enable resolution-agnostic upsampling.

Method

LiDAR SR typically converts sparse 3D point clouds to 2D range images, processes them with deep learning networks to predict high-resolution range images, and then back-projects to dense 3D point clouds.

In practice

• Use range images for efficient LiDAR data processing.
• Employ model-based unrolling for resource-constrained systems.
• Consider implicit functions for flexible resolution output.

Topics

• LiDAR Super-Resolution
• Autonomous Driving
• Deep Unrolling Networks
• Implicit Neural Representations
• Transformer and Mamba Models

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, AI Researcher, Deep Learning Engineer, Robotics Engineer

Related on AIssential

• Towards Compact Autonomous Driving Perception with Balanced Learning and Multi-sensor Fusion — The model integrates multi-sensor data and multi-task learning into a compact architecture for efficient autonomous driving perception.
• MapSR: Prompt-Driven Land Cover Map Super-Resolution via Vision Foundation Models — MapSR uses prompt-driven vision foundation models for efficient, training-free land-cover super-resolution with minimal annotation.
• OmniLiDAR: A Unified Diffusion Framework for Multi-Domain 3D LiDAR Generation — OmniLiDAR unifies 3D LiDAR generation across diverse sensing domains using a text-conditioned diffusion framework.
• GEM: Generating LiDAR World Model via Deformable Mamba — GEM is a LiDAR world model using deformable Mamba to improve autonomous driving simulation by addressing point cloud disorder and object distinction.
• teasr: training-efficient any-step diffusion transformer for real-world image super-resolution — TEASR uses self-adversarial distillation within a single diffusion model for efficient, flexible Real-ISR, eliminating auxiliary teachers.
• EfficientPENet: Real-Time Depth Completion from Sparse LiDAR via Lightweight Multi-Modal Fusion — EfficientPENet achieves real-time depth completion on edge devices by combining a lightweight ConvNeXt backbone with sparsity-invariant convolutions and CSPN refinement.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.