SteerMap

Long-Horizon Dense Mapping

SteerMap

Steering Heavyweight AI with Lightweight AI for Long-Horizon Dense Mapping

SteerMap turns the lightweight mapping AI already running for localization into a geometric control plane for an on-demand vision foundation model. It decides what to preserve inside each call and where dense local maps belong across calls.

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

Problem and insight

Long-horizon mapping changes the task.

A kilometer-scale route produces thousands of images, exceeds the memory budget of a single VFM call, and revisits regions whose poses may later change after loop closure.

Challenge 01

Global-geometry inconsistency

Independent VFM calls produce plausible local maps, but later trajectory corrections do not automatically repair maps built earlier.

Challenge 02

Computational inefficiency

Cross-frame attention repeatedly processes overlapping visual evidence, while the number of expensive calls grows with trajectory length.

Keep lightweight AI in the loop; invoke heavyweight AI on demand.

Unlike existing pipelines that ask a heavyweight VFM to both reconstruct local geometry and maintain global consistency, SteerMap uses persistent lightweight geometry to steer when and how the VFM is used.

Figure 5 comparing heavyweight AI in the loop with SteerMap's lightweight AI in the loop and heavyweight AI on demand paradigm
Figure 5 from the paper: long-horizon dense mapping pipelines.

One geometry stream, two levels of steering.

The always-on mapper continuously exposes keyframes, camera poses, sparse map points, and multi-view observations. The same state reduces computation inside the VFM and consolidates dense maps outside it.

Figure 6 showing the SteerMap system architecture
Figure 6 from the paper: SteerMap system architecture.

Method

Steer within each call. Orchestrate across calls.

Inside-Model SteeringSpend attention on geometrically important evidence while compressing repeated content.

Outside-Model OrchestrationReuse valid dense geometry as poses evolve and reconstruct only what can no longer be explained coherently.

Inside-Model Steering

Geometric Saliency-Guided Token Compression

Project sparse multi-view observations onto the token grid, protect reconstruction-critical evidence, merge redundant tokens before cross-frame global attention, and restore the full layout before downstream prediction.

  • Cross-view support
  • Geometric novelty
  • Surface boundaries
Figure 7 showing geometric saliency-guided token compression and local map generation
Figure 7 from the paper: inside-model steering and local map generation.
Outside-Model Orchestration

Geometric Consistency-Guided Global Map Consolidation

Use the latest global trajectory to reanchor reusable local maps, selectively reconstruct inconsistent frame sets, and refine dense confidence through sparse–dense geometric agreement.

  • Local-map reanchoring
  • Selective reconstruction
  • Spatial refinement
Figure 10 showing geometric consistency-guided global map consolidation
Figure 10 from the paper: outside-model orchestration and global map generation.

Videos

Long-horizon mapping in motion.

These demonstrations complement the paper with causal map evolution and full-sequence comparisons. Open full screen to inspect trajectory changes 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 · Field study

Outdoor dataset comparison

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

Demo 03 · KITTI 00

Comparison with long-horizon baselines

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

Review artifact

Inspect the pipeline.

The anonymous artifact contains the runnable KITTI 00 pipeline, causal geometric-state reader, configurations, evaluation tools, and instructions for obtaining the review data.

Open anonymous code ↗