Compressing Spatial Maps into Semantic Primitives

59 concepts53 papers2 proposals

Robots and autonomous vehicles rely heavily on high-definition (HD) maps to navigate. Traditionally, these maps are built as dense 3D point clouds, volumetric occupancy grids, or high-resolution meshes. While geometrically rich, these representations carry an immense "data tax." A single high-fidelity LiDAR scan or 3D mesh can require megabytes of storage. Scaling this to city-wide or global levels creates a massive bottleneck: maps become too large to store entirely in system memory, too heavy to stream over standard wireless networks, and too bandwidth-intensive to share in multi-agent systems.

However, a fundamental insight has emerged across recent robotics research: high-fidelity geometric data is largely redundant for navigation. A self-driving car does not need to know the exact millimeter-level surface roughness of a curb to stay in its lane; it only needs the mathematical boundary of the lane. A drone does not need a fully textured 3D mesh of a room to localize; it needs to know where the walls, doors, and large furniture are.

By shifting the fundamental unit of mapping from raw, low-level geometric elements (points, voxels, surfels) to sparse, high-level semantic primitives, researchers are achieving map compression ratios of three to four orders of magnitude. This paradigm allows robots to maintain centimeter-level localization accuracy while reducing map sizes from hundreds of megabytes to a few kilobytes.

Geometric Abstraction via Parametric Primitives

The most direct way to compress spatial data is to replace thousands of individual 3D points with a single mathematical equation representing a semantic landmark. For instance, a long stretch of a highway lane or a building facade can be abstracted into a line, a spline, or a plane.

In autonomous driving, Wei-Chiu Ma et al. demonstrated in Exploiting Sparse Semantic HD Maps for Self-Driving Vehicle Localization that representing HD maps using sparse geometric primitives—such as polylines for lane graphs and point coordinates for landmarks—reduces map data requirements by over 99% while maintaining centimeter-level lateral accuracy. Similarly, Kun Jiang et al. in Monocular Vehicle Self-localization method based on Compact Semantic Map vectorized environmental landmarks into lines for lane markings and points for traffic signs, showing that sub-meter accuracy is achievable using extremely lightweight data structures.

This abstraction applies equally well to 3D structures. Instead of dense point clouds, the authors of Multi-Session, Localization-oriented and Lightweight LiDAR Mapping Using Semantic Lines and Planes abstracted complex surfaces into minimal geometric primitives like infinite lines for poles and planes for facades. This filters out dynamic noise and reduces storage over a thousandfold. For more complex curved objects, Weikun Zhen et al. in Unified Representation of Geometric Primitives for Graph-SLAM Optimization Using Decomposed Quadrics and Andrew Davison et al. in Simultaneous Localisation and Mapping with Quadric Surfaces utilized decomposed quadrics (such as spheres, cylinders, and cones). By storing only a few parameters (pose, scale, and shape type) rather than thousands of surface points, the map retains its structural and semantic integrity at a fraction of the computational cost.

Object-Centric Compression and Centroid Mapping

Beyond fitting lines and planes to surfaces, another powerful mechanism treats the environment as a sparse collection of discrete, recognizable objects. If a robot can recognize an object instance, it can replace the dense geometry of that object with a single semantic label and a spatial anchor (like a 3D bounding box or a centroid coordinate).

This concept was pioneered by Renato F. Salas-Moreno et al. in SLAM++: Simultaneous Localisation and Mapping at the Level of Objects, which stored a single high-quality 3D model for repeated object instances and recorded only the 6-DoF pose for each instance, achieving compression ratios as high as 1/70. In outdoor environments, Georgi Pramatarov et al. introduced BoxGraph: Semantic Place Recognition and Pose Estimation from 3D LiDAR, which encodes object shapes as simple 3D bounding boxes to reduce map sizes by 25 times compared to existing semantic methods.

Taking this to the extreme, the authors of That's My Point: Compact Object-centric LiDAR Pose Estimation for Large-scale Outdoor Localisation demonstrated that a 1.41 MB raw point cloud can be compressed down to just 1.33 kB by storing only the 3D coordinates and semantic classes of static entities like buildings or poles. Because the spatial arrangement of these centroids provides unique geometric constraints, the robot can achieve centimeter-level localization without needing any raw geometric points. A similar sparse strategy is utilized in Key-Region-Based UAV Visual Navigation, where mapping a small number of distinct semantic objects (like cars) instead of thousands of fragile keypoints enables extreme map compression while maintaining robust localization.

Topological Abstraction via Semantic Scene Graphs

When maps are reduced to discrete objects or primitives, the spatial and functional relationships between these entities can be formalized as a Semantic Scene Graph. In this representation, nodes correspond to objects or geometric primitives, and edges represent their spatial or topological connections.

Scene graphs provide a highly compressed, non-linear representation that accelerates matching and drastically cuts down communication bandwidth. For example, Chuhao Liu et al. developed SG-Reg: Generalizable and Efficient Scene Graph Registration, showing that representing 3D environments as semantic scene graphs reduces communication bandwidth requirements by over 25 times compared to traditional visual loop closure methods. This enables multi-agent map merging in bandwidth-constrained robotics applications.

For localization and retrieval, Francis Engelmann et al. introduced SceneGraphLoc: Cross-Modal Coarse Visual Localization on 3D Scene Graphs, which distills multi-modal information into fixed-size object embeddings within a scene graph. This abstraction reduces storage requirements by three orders of magnitude compared to traditional image-based descriptor databases. Similarly, Daniel Barath et al. in "Where am I?" Scene Retrieval with Language leveraged semantic graphs to reduce storage to a few hundred kilobytes, enabling millisecond-level language-based scene retrieval for privacy-sensitive or low-bandwidth embodied AI applications.

Task-Aware and Prior-Driven Data Reduction

Why store a dense map at all if we can leverage existing, lightweight global priors or optimize the map based on the robot's specific task?

One approach is to match real-time sensor data directly against sparse, crowdsourced 2D maps like OpenStreetMap (OSM). Paul-Edouard Sarlin et al. developed OrienterNet: Visual Localization in 2D Public Maps with Neural Matching, which matches a neural Bird's-Eye View derived from a query image against a 2D semantic map. This reduces map storage requirements by approximately 8,800 times compared to traditional 3D point clouds. Furthermore, Junhui Wang et al. showed in OpenBench: A New Benchmark and Baseline for Semantic Navigation in Smart Logistics that lightweight, crowdsourced map data combined with foundation models can completely replace expensive high-precision maps by leveraging visual reasoning and common-sense interpretation.

Alternatively, if a map must be built from scratch, information theory can dictate what to keep. Jaein Lim proposed an optimization framework in Information-theoretic Abstraction of Semantic Octree Models for Integrated Perception and Planning that allows for non-uniform map resolution. By preserving task-relevant semantic categories (like roads) in high detail while aggressively pruning irrelevant categories (like vegetation or sky), the overall data footprint is minimized without sacrificing the fidelity critical for autonomous navigation.

Key Foundational Work

Localizing while Mapping: A Segment Approach (2002) — contributed 2 concept elements to this direction.
Global Symbolic Maps from Local Navigation (1991) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 49 papers
  2. Version 1 · 2026-10-04 · 44 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_map_compression_49cce521,
  title = {semantic map compression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-map-compression/rdr_49cce52157f4466ab916d9c6512c6c63},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_semantic_map_compression_49cce521_v2,
  title = {semantic map compression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-map-compression/rdr_49cce52157f4466ab916d9c6512c6c63/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake