Skip to content
DepthWizardDepthWizard

Research paper

Fine-Tuning a Satellite-Pretrained DINOv3 for Single-View Metric Height Estimation from PNG/JPEG and Georeferenced Imagery Abhay Kale, Akash Chaudhari, Somesh Padsalge. Department of AI & Data Science, Jawaharlal Nehru Engineering College, Chhatrapati Sambhaji Nagar.

Summary

Section titled “Summary”

The paper fine-tunes a DINOv3 ViT-L/16 encoder, pretrained on 493 M satellite images, to predict metric nDSM at a canonical 0.5 m GSD. A DPT decoder feeds regression, adaptive-bin and semantic heads, joined by a learned gate and trained with a height-stratum-balanced loss. The best configuration reaches a validation RMSE of 2.61 m (r = 0.92) on GAMUS. A later configuration scores 3.32 m on the 2,861-tile GAMUS test split and 4.97 m on out-of-domain DFC23. The paper also shows that 85.5 % of a later regression came from the 0–2 m band, explains why adding a DEM to a predicted nDSM double counts buildings, and describes the ONNX export and browser-based 3D viewer.

Structure

Section titled “Structure”
  1. Introduction
  2. Related work: monocular depth estimation; height from remote sensing; foundation encoders and datasets
  3. Method: problem formulation, architecture, objective, optimisation, georeferenced calibration, inference
  4. Experimental setup: datasets, model versions, metrics
  5. Results: validation across versions, error by height stratum, held-out and cross-domain evaluation, absolute DSM recovery, qualitative results
  6. Analysis of design findings: padding leakage, cross-sensor radiometry, encoder unfreezing, stratum weighting, DEM double counting
  7. System and visualisation
  8. Discussion and limitations
  9. Conclusion and future work

These docs expand on every section. See Architecture, Training, Benchmarks and Design findings.

Section titled “Related work compared”
Family Works
Monocular depth Eigen et al. (SiLog), AdaBins, DPT, Depth Anything V1/V2, Marigold
Height from remote sensing IM2HEIGHT, IM2ELEVATION, HTC-DC Net
Foundation encoders ViT, DINOv2, DINOv3
Datasets & DEMs GAMUS, SynRS3D, DFC23, SRTM, Copernicus DEM

Cite

Section titled “Cite”
@misc{kale2026depthwizard,
  title  = {Fine-Tuning a Satellite-Pretrained {DINOv3} for Single-View Metric Height
            Estimation from {PNG/JPEG} and Georeferenced Imagery},
  author = {Kale, Abhay and Chaudhari, Akash and Padsalge, Somesh},
  year   = {2026},
  note   = {Department of AI \& Data Science, Jawaharlal Nehru Engineering College},
}