SteerMap

Long-horizon streaming dense mapping

SteerMap

Orchestrating Vision Foundation Models with Lightweight Geometry for Long-Horizon Dense Mapping

SteerMap keeps lightweight mapping AI online and invokes a heavyweight vision foundation model on demand. Geometry steers what the model preserves, where local maps move, and which inconsistent regions must be reconstructed.

Watch videos Anonymous code ↗
Lightweight AI in the loop Heavyweight AI on demand

Overview

A different pipeline for the long horizon.

Existing systemskeep a heavyweight VFM in the loop for both dense reconstruction and global consistency.

SteerMapkeeps lightweight geometry in the loop and invokes the VFM only when dense reconstruction is needed.

Paper Figure 1 comparing the existing heavyweight-AI-in-the-loop pipeline with SteerMap's lightweight-AI-in-the-loop pipeline
Figure 1 from the paper: long-horizon dense mapping pipelines.

One geometry stream steers both local and global mapping.

The always-on geometry mapper continuously exposes camera poses, sparse map points, and observations. SteerMap uses this state to accelerate local-map generation and maintain global consistency as maps evolve.

Paper Figure 5 showing the SteerMap system architecture
Figure 5 from the paper: SteerMap system architecture.

Method

Two levels of geometric steering.

Inside the model — Faster local mapping.Geometry-aware token compression accelerates VFM inference while preserving local-map quality.

Outside the model — Globally consistent mapping.Reanchor valid maps and reconstruct only inconsistent regions as poses evolve.

Inside-model steering

Geometric Saliency-Guided Token Compression

Protect geometrically supported tokens, merge redundant tokens before cross-frame attention, and restore the full layout for prediction.

Paper Figure 6 showing geometric saliency-guided token compression and local-map generation
Figure 6 from the paper.
Outside-model orchestration

Geometric Consistency-Guided Global Map Consolidation

Reanchor reusable local maps, selectively reconstruct inconsistent segments, and refine dense confidence with reliable sparse geometry.

Paper Figure 9 showing geometric consistency-guided global map consolidation
Figure 9 from the paper.

Videos

Long-horizon mapping in motion.

All videos are presented with native browser controls. Open full screen to inspect trajectory evolution and reconstruction details.

Demo 01 · KITTI 00

First-person streaming reconstruction

A 30-second view of SteerMap processing KITTI 00 as the vehicle traverses the sequence.

Demo 02 · KITTI 00

Comparison with long-horizon baselines

Four-method comparison on KITTI 00 (3.724 km).

Demo 03 · Field study

Outdoor dataset comparison

Four-method comparison on our outdoor traversal (1.17 km).

Quantitative results

Accuracy at the long horizon, without the long wait.

Selected quantitative results below are extracted directly from the submitted paper. Click any figure to inspect it at full resolution.

CDFs of accuracy, completeness, and absolute trajectory error on the 1.17-kilometer outdoor field study
Field-study error distributions. SteerMap shifts accuracy, completeness, and absolute trajectory error toward the low-error region throughout the 1.17-km outdoor traversal.
Quantitative comparison on the 1.17-kilometer outdoor field study
Outdoor field-study results. SteerMap jointly improves trajectory accuracy, reconstruction quality, coverage, and end-to-end mapping time.
Long-horizon mapping comparison on KITTI sequences 00, 02, and 08
Long-horizon KITTI results. Results on the three longest evaluated KITTI sequences, covering different loop directions and more than 3 km per sequence.
SteerMap component, local-map generation, and global-map consolidation efficiency breakdown
Efficiency and quality breakdown. The ablations expose where computation is saved and how selective global-map consolidation approaches always-reconstruct quality with fewer VFM calls.

Code and data

Anonymous review artifact.

The repository contains the runnable KITTI 00 pipeline, causal SLAM revision reader, configuration, evaluation tools, and instructions for obtaining the review data.

Open anonymous code ↗