Dilavesh/altimap-demo

AltiMap prepared demo scenes

Seven complete saved outputs of the AltiMap single-image elevation pipeline, prepared on 2026-10-05. The catalog opens with Sikkim satellite imagery, followed by Pittsburgh bridges and three reference-scored examples. Philadelphia City Hall is excluded from this demo. These support the Vercel viewer while its live GPU inference VM is offline. This repository stores files; it does not run inference. Buildings use the image-derived textured renderer. Map-assisted building reconstruction has been reverted and is not included.

index.json lists scenes; each result.json contains display geometry, metric elevation grids, model provenance, facility positions and the original validation. PNGs hold RGB/height/classes, terrain and signed-error previews. Original nDSM/DSM GeoTIFFs and GeoJSON exports accompany them. Processing was Fast (one pass), with SRTM for georeferenced scenes and no GCP correction.

Scene	Reference	All-valid-pixel RMSE (m)
GAMUS DC_04_27 mixed urban tile	LiDAR-derived AGL/nDSM	2.8279
Chevy Chase suburb	USGS LiDAR absolute DSM	4.2845
Pittsburgh hills	USGS LiDAR absolute DSM	6.2394
Pittsburgh bridges	No independent reference attached	Unvalidated
Chungthang, Sikkim town	No independent reference attached	Unvalidated
Namchi, Sikkim town	No independent reference attached	Unvalidated
Chungthang, Sikkim forest	No independent reference attached	Unvalidated

The Sikkim and bridges scenes have no attached independent height reference; its mapped bridge positions are OSM context, not proof of measured deck heights. These are scene-specific comparisons, not independent survey-grade reconstruction. Building meshes are approximate visualizations. Reference/imagery dates and vertical datums can differ: NAIP reference rasters use NAVD88, whereas the SRTM-based outputs use EGM96. Published metrics retain those differences. Raster probes and validation use original model outputs.

Sources and attribution

• AltiMap source, including setup, architecture and evaluation reports.
• AltiMap model weights: v1 with 75% v2 building blend.
• GAMUS, CC BY 4.0: GAMUS: A Geometry-aware Multi-modal Semantic Segmentation Benchmark for Remote Sensing Data, arXiv:2305.14914. The imagery preview is derived from GAMUS and the prediction/reference comparison uses its corresponding AGL layer.
• Sikkim imagery: © Maxar Open Data Program, WorldView-2 (Chungthang, 2022-03-07) and WorldView-3 (Namchi, 2022-03-14), North India Floods event imagery, CC BY-NC 4.0. RGB previews are derived crops; these are pre-flood acquisitions, not current disaster imagery. The exported grid spacing is 0.305 m and should not be confused with the sensor's native resolution (approximately 0.49 m / 0.37 m). Building and canopy heights are AltiMap model predictions, not Maxar measurements.
• NAIP aerial imagery: USDA, obtained through Microsoft Planetary Computer. Absolute-height references: USGS 3DEP airborne LiDAR. Original uploads and reference rasters are not included.
• Facility/bridge/flood context: © OpenStreetMap contributors, ODbL. Positions were cached at processing time and may be incomplete or outdated.
• Forest canopy specialist: Meta CHMv2 / DINOv3, public model mirror and license. Built with DINOv3.
• RS3DAda / SynRS3D model architecture and DINOv2 encoder: see the project's third-party notices.

Use this bundle as a non-commercial research/demo artifact, retaining the above credits and following each upstream source's terms. A single blanket license does not replace the source licenses. The v2 model includes SynRS3D training data, whose non-commercial terms are recorded in the AltiMap model card. No checkpoint weights are duplicated in this dataset.