ExelMap: Element-Based HD Map Change Detection and Update

Last updated: 2026-05-09

Why It Matters

ExelMap reframes HD map maintenance from "is this frame inconsistent with the map?" to "which map element changed, and should the element be inserted, deleted, or preserved?" That distinction is operationally important because fleet map updates need reviewable evidence, not only a scalar anomaly score.

For airside autonomy, the same pattern applies to stand lead-in lines, stop bars, service-road markings, temporary construction geometry, barrier placements, and docking landmarks. A vehicle can detect a discrepancy, but map operations need the exact element, change type, persistence evidence, and rollback impact before publishing a map tile.

Paper Snapshot

Item	ExelMap position	Airside interpretation
Core task	Explainable element-based HD map change detection and update	Identify the changed stand, marking, lanelet, barrier, or map feature rather than only flagging the route
Inputs	360-degree camera features plus stale HD map prior	Sensor evidence plus active airport map tile
Backbone	LaneSegNet-style vector map generation	Compatible with lanelet/segment representations, but airport semantics need custom classes
Prior encoding	Geometry and semantic attributes from the stale map	Use prior linework, markings, and feature type as a second input
Change heads	Separate element-wise insertion and deletion heads	Produce reviewable "add this feature" and "remove this feature" proposals
Output	Updated local HD map plus change map	Candidate map patch plus audit trail
Demonstration	Argoverse 2 Map Change/TbV pedestrian-crossing changes	Road-crossing evidence, not direct apron validation

What It Adds

Problem with prior work	ExelMap response	Practical value
Frame-level change labels do not say what changed	Element-level change status	Human reviewers can inspect specific elements
Standard mAP hides whether a model updated the stale region	Separate change-detection and updated-map evaluation views	Avoids inflated results from simply copying most unchanged prior elements
Prior-aided generation can silently pass stale map content through	Explicit change heads beside generation heads	Makes the "trusted prior vs current sensor" decision visible
Synthetic changes may not generalize to real changes	Evaluation on real Argoverse 2 map changes	Useful warning for airside synthetic-change generators
Single-frame evidence is unstable under occlusion	Paper identifies need for true multi-frame handling	Airside deployment should aggregate across passes before map publication

Evaluation Checklist

Evaluation question	Metric family	Airside acceptance use
Did the system detect that any change exists?	Type-agnostic change/no-change accuracy	Alert routing and tile quarantine
Did it detect the right change type?	Type-aware insertion/deletion accuracy	Decide whether to add, delete, or keep a feature
Did it localize the changed element?	Changed-element IoU/localization accuracy	Send a precise patch to map QA
Is the updated map geometrically good?	LaneSegNet-style AP on updated map	Confirm planner/localizer quality after patch
Is unchanged prior preserved?	Unchanged-element preservation	Prevent map erosion from noisy perception
Does evidence persist over time?	Multi-frame or multi-pass aggregation	Required before publishing an airside static-layer update

Airside Translation

Airside element	Candidate change	Default policy
Stand lead-in/lead-out marking	Inserted, repainted, shifted, removed	Quarantine tile and require visual or survey review
Service-road lanelet	Geometry or lane-count change	Require multi-pass evidence and route-regression test
Stop bar / hold line	Inserted, deleted, shifted	Human approval; never auto-delete from production map
Temporary barrier / cone line	Inserted obstacle or work-zone edge	Live restriction first; static map update only if persistent and approved
GSE parking outline	Repainted or newly observed marking	Low-risk update if repeated and not conflicting with airport ops data
Aircraft/GSE object	Apparent deletion/insertion in point cloud	Do not update permanent map; route to movable-static or dynamic layer

Implementation Guidance

1. Keep the ExelMap-style "change map" separate from the updated map. The change map is the review artifact; the updated map is only a candidate product.
2. Extend the change vocabulary beyond insertion/deletion before airside use: geometry edit, marking type, regulatory status, topology, movable-static, and unknown.
3. Do not rely on standard map-generation mAP alone. Report update quality separately for changed and unchanged elements.
4. Require temporal evidence for any deletion. A missing marking can be occluded by aircraft, wet surface glare, snow, service vehicles, or camera exposure.
5. Apply stricter thresholds near aircraft stands, hold points, terminal edges, and pedestrian crossings than on open service roads.
6. Store original prior element ID, proposed replacement geometry, sensor frames, timestamp, vehicle pose quality, and reviewer decision with every candidate patch.
7. Use ExelMap as a design pattern, not a deployable airside model: its public experiment focuses on pedestrian crossings in road scenes.

Fit With Existing Map Stack

Existing repo concept	ExelMap connection
moved-object-and-map-change-datasets.md	ExelMap is the element-level complement to scenario/frame-level map-change datasets
hd-map-change-detection-maintenance.md	Supplies the detection/update primitive needed by fleet map maintenance
map-tile-versioning-distribution.md	Candidate element patches should become tile diffs only after QA and localization regression
lt-mapper-khronos-lifelong-mapping.md	Multi-session lifelong evidence should decide promotion, not a single ExelMap frame

Sources

• ExelMap arXiv abstract and paper: https://arxiv.org/abs/2409.10178
• ExelMap HTML version: https://ar5iv.org/html/2409.10178v1
• Argoverse 2 Map Change Dataset overview: https://www.argoverse.org/av2.html
• Argoverse TbV user guide: https://argoverse.github.io/user-guide/datasets/map_change_detection.html
• Trust, but Verify paper: https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/2b8a61594b1f4c4db0902a8a395ced93-Paper-round2.pdf
• Local context: Moved-Object and Map-Change Datasets
• Local context: HD Map Change Detection, Maintenance, and Fleet-Based Updates