Research map / Point clouds / Point cloud classification

DeepPointMap: Advancing LiDAR SLAM with Unified Neural Descriptors

AAAI 2024 · Point clouds › Point cloud classification · approach: LiDAR point cloud compression and completion · 7 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DeepPointMap, a learning-based LiDAR SLAM framework that encodes point clouds into sparse unified neural descriptors (PointNeXt encoder + transformer/similarity/offset/overlap decoder) used for odometry, loop closure, and multi-agent cooperative mapping.

Problem

LiDAR SLAM faces a conflict between high localization accuracy, which needs dense point clouds, and memory-efficient maps, which need compact descriptors; prior geometric and learned methods also handle only single-scale registration and are not unified across odometry and loop closure.

Approach

Extract sparse neural descriptors with a PointNeXt encoder and use a single descriptor-wise transformer decoder with similarity, offset (weighted SVD) and overlap heads to perform multi-scale registration (scan-to-scan, scan-to-map, map-to-map) directly on the descriptors, trained with pairing, coarse pairing, and offset losses plus curriculum learning and random occlusion augmentation.

Key results

• On the 12 test sequences of SemanticKITTI, KITTI-360 and MulRan, DeepPointMap attains the best APE on 8 sequences in the localization accuracy table (e.g., KITTI-360 08: 5.26 vs. 8.24 for the next best; MulRan Riverside02: 17.48 vs. 27.16 for KISS-ICP).
• The map representation saves up to 70% memory on SemanticKITTI 07 and about 50% on KITTI-360 08 compared to original point cloud, voxel hashmap and mesh.
• Ablation on SemanticKITTI 08-10 shows removing curriculum learning and random occlusion raises APE drastically (e.g., 174.57 vs. 4.7 on sequence 08).

Measured effects reported in the paper

Surprising findings

• GeoTransformer, a point cloud registration method, yields very large trajectory errors (e.g., over 600 APE on KITTI-360 07) when used for scan-to-scan trajectory estimation.
• Removing the offset head causes a catastrophic failure on SemanticKITTI 10 (63.26 vs. 1.35 baseline).

Limitations

• Requires precise ground-truth pose labels for each LiDAR scan for training, which are often of low quality in driving datasets.
• Weaker performance in sparse distant countryside sequences (e.g., KITTI-360 07, 10) lacking reference objects and geometric information.
• Only LiDAR modality; plans to incorporate vision in future work.

Experimental setup

Datasets: SemanticKITTI KITTI-360 MulRan KITTI-Carla

Benchmarks: SemanticKITTI KITTI-360 MulRan

Metrics: APE memory size

Baselines: LeGO-LOAM SC-LeGO-LOAM MULLS CT-ICP GeoTransformer KISS-ICP Voxel Hashmap Mesh

Builds on: PointNeXt GeoTransformer InfoNCE weighted SVD

Closest papers by meaning

• DeepMapping: Unsupervised Map Estimation From Multiple Point Clouds (CVPR 2019 · similarity 0.93)
• DeepMapping2: Self-Supervised Large-Scale LiDAR Map Optimization (CVPR 2023 · similarity 0.89)
• 3D LiDAR Mapping in Dynamic Environments using a 4D Implicit Neural Representation (CVPR 2024 · similarity 0.86)
• MAGiC-SLAM: Multi-Agent Gaussian Globally Consistent SLAM (CVPR 2025 · similarity 0.86)
• LightLoc: Learning Outdoor LiDAR Localization at Light Speed (CVPR 2025 · similarity 0.85)
• LiSA: LiDAR Localization with Semantic Awareness (CVPR 2024 · similarity 0.85)