AltiMap height model

Single-image height estimation for aerial and satellite RGB imagery: for every pixel, the height above ground in metres (nDSM) plus an 8-class land-cover map. Built for Smart India Hackathon 2026, problem statement 26175 (ISRO, "DepthWizard"), as the model behind AltiMap, which turns one image into metric elevation GeoTIFFs and a 3D city model. Setup and usage: the repo's SETUP.md.

Model

• Architecture: RS3DAda: DINOv2 ViT-L/14 encoder + DPT decoder with a height-regression head and a segmentation head (SynRS3D, NeurIPS 2024). Needs the model code from JTRNEO/SynRS3D.
• Starting point: the authors' RS3DAda height weights.
• Fine-tuning (best.pth, v1): on GAMUS (0.33 m aerial RGB with airborne-LiDAR nDSM and land-cover labels over Washington DC, Philadelphia and New York), BitFit (encoder biases) + decoder and heads, loss L1 + 0.5 x gradient L1 + 0.2 x cross-entropy, 518 px crops, bf16.
• v2 (v2/best.pth): continues from v1 with the last 8 encoder blocks unfrozen, a height-weighted L1 (a pixel at h metres counts 1 + h/10 times), 30 % SynRS3D samples (high-rises, hills) and 50 % satellite-style degradation (1.3–6× coarser, blur, haze). 6.5 h on an RTX A5000, best checkpoint at step 25,500.

Results (GAMUS test split, 500 tiles, 4-flip test-time averaging)

How AltiMap uses them

v2 wins on GAMUS but generalises worse to other imagery: against USGS airborne LiDAR on NAIP scenes it over-reads trees and ground. On buildings, though, it is right where v1 isn't. So the app combines three models by land cover:

• v1 gives heights everywhere.
• Building pixels take the mean of v1 and v2. Building RMSE: city 44.5 → 41.9 m, hills 5.2 → 4.4 m, GAMUS val 2.57 → 2.40 m. The city's tallest objects rise from 66 to 104 m (LiDAR 151 m).
• Tree pixels in extensive forest (≥ 80 % canopy within 150 m) come from Meta's CHMv2 canopy model.

Limitations (measured, not guessed)

• Out of domain, by landscape (NAIP aerial RGB vs USGS 3DEP airborne LiDAR, one pass, the app's three-model pipeline; v1 alone in brackets). Absolute DSM = these heights on Copernicus GLO-30, DEM-consistent:

  Scene	nDSM RMSE (predict 0)	nDSM bias	DSM RMSE (GLO-30 alone)
  Dense city (0.3 m)	30.1 m (42.4) [29.8]	−1.2 m [−6.8]	35.0 m (36.3) [34.7]
  Suburb (0.6 m)	4.5 m (6.5) [4.5]	+1.4 m [+1.3]	3.97 m (4.02) [3.99]
  Hilly town (0.6 m)	3.5 m (6.5) [3.7]	−0.4 m [−0.7]	5.02 m (5.53) [5.04]
  Forest (0.6 m)	15.7 m (24.7) [18.5]	−10.8 m [−15.9]	8.75 m (8.53) [9.03]

  Very tall towers still read low (104 m vs 151 m), and forest canopy reads ~11 m low. Object heights hold up to ~1–2 m pixels and fade to flat by 5–10 m. ISRO's evaluation imagery is 0.6 m Cartosat-2S over India, outside the training domain (aerial imagery of US cities).

• Heights are above ground. Absolute elevation needs a terrain model; AltiMap adds Copernicus GLO-30 ground for georeferenced inputs.

Usage

from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("Dilavesh/altimap-height", "best.pth")  # public, no login needed

# with the AltiMap repo checked out and SynRS3D cloned into viewer/cache/SynRS3D:
from pathlib import Path
from viewer.height_model import load_model, predict
model, device = load_model(Path(ckpt), Path("viewer/cache/SynRS3D"))
ndsm_m, oem_classes = predict(model, rgb_uint8_hxwx3, device, tta=True)  # input at ~0.33 m/px

code/ holds the training and data scripts used for these weights.

License and attribution

Weights: MIT, like the RS3DAda base model. Please credit:

• RS3DAda / SynRS3D: Song et al., SynRS3D: A Synthetic Dataset for Global 3D Semantic Understanding from Monocular Remote Sensing Imagery, NeurIPS 2024 (code and base weights MIT).
• GAMUS (CC BY 4.0): GAMUS: A Geometry-aware Multi-modal Semantic Segmentation Benchmark for Remote Sensing Data, arXiv:2305.14914.
• DINOv2 (Apache 2.0), Oquab et al., Meta AI.

The v2 weights were also trained on SynRS3D data, which is CC BY-NC 4.0: treat v2 as non-commercial.