Technical Report for DepthWizard

Technical overview

Model architectureExplains the satellite-pretrained DINOv3 encoder, DPT decoder, and multi-head height prediction. See how regression, adaptive depth bins, and land-cover segmentation combine into one estimate.

Training & fine-tuningCovers the datasets, preprocessing, augmentations, and fine-tuning recipe, then follows the model through its five versions. Highlights the choices that improved results across sensors and regions.

Evaluation & resultsPresents held-out GAMUS and cross-domain DFC23 results alongside LiDAR and Cartosat validation. Error breakdowns by height range and landscape show where predictions hold up.

Georeferencing & scaleExplains how plain images get relative scale and how GeoTIFF metadata anchors predictions geographically. Shows how DEM correction produces an absolute DSM without counting buildings twice.

System architectureTraces the path from the web interface through hosted or self-hosted inference to 3D visualization and exports. Includes the API, deployment options, and ONNX support.