DepthWizard

Single-view satellite imagery to metric elevation and immersive 3D terrain

An end-to-end geospatial reconstruction system that transforms georeferenced and non-georeferenced optical imagery into elevation products and navigable 3D scenes—without LiDAR input or OSM building geometry.

---

Important links

Resource	Link
🚀 Live demo	Launch DepthWizard
🎬 Demo video	Watch the complete walkthrough
📘 Technical documentation	Read the technical report
💻 Source code	View the GitHub repository

---

Results

DepthWizard selects an appropriate reconstruction path for urban, mountainous, forested, and sparse scenes while preserving the source image as the visual texture. The following comparisons show the original 2D optical input beside the generated 3D scene.

🏙️ Urban reconstruction

2D optical input	DepthWizard 3D reconstruction
New York City · Dense urban scene	LoD2 buildings with orthophoto roofs and generated facade materials
Rotterdam · Mixed urban morphology	Geometrically regularized building reconstruction
Delft Old Town · Compact historic fabric	Textured roofs with footprint-aware extrusion

⛰️ Mountain and natural terrain

2D optical input	DepthWizard 3D reconstruction
Atlas Mountains	Elevation-displaced, optical-textured terrain
Manali	High-relief terrain reconstruction
Wengen	Continuous mountain surface with optical projection

🌲 Forest and canopy reconstruction

2D optical input	DepthWizard 3D reconstruction
Western Ghats · Dense vegetation	Terrain-following canopy and forest assets
Arunachal Pradesh · Forested relief	Forest-aware presentation with preserved terrain structure

🌾 Sparse vegetation

2D optical input	DepthWizard 3D reconstruction
Thar Desert · Sparse settlement and vegetation	Sparse vegetation assets placed without modifying terrain or buildings

---

Benchmarking

DepthWizard evaluates generated absolute Digital Surface Models against OpenTopography COP30 reference elevations. The comparison pipeline aligns the reference data to the optical grid and reports RMSE, MAE, Pearson correlation, and the project accuracy score across four landscape typologies.

⛰️ Hilly and mountainous regions

Location	Estimated height	Reference height	RMSE ↓	MAE ↓	Pearson ↑	Accuracy ↑
Wengen Slope, Switzerland	1294.41 m	1294.41 m	1.19 m	0.34 m	1.0000	95.72%
Interlaken Hills, Switzerland	1980.12 m	1984.18 m	0.96 m	0.36 m	1.0000	96.17%
Atlas Mountains, Morocco	3159.70 m	3158.06 m	3.03 m	1.72 m	1.0000	73.14%

🏙️ Urban areas

Location	Estimated height	Reference height	RMSE ↓	MAE ↓	Pearson ↑	Accuracy ↑
Rotterdam Centre, Netherlands	3.48 m	3.47 m	1.03 m	1.72 m	0.95	99.71%
Delft Old Town, Netherlands	7.26 m	5.46 m	1.03 m	1.06 m	0.89	65.78%
New York City, USA	174.57 m	174.32 m	1.01 m	1.56 m	0.90	90.89%

🌾 Sparse regions

Location	Estimated height	Reference height	RMSE ↓	MAE ↓	Pearson ↑	Accuracy ↑
Rajasthan (Thar Desert), India	10.57 m	10.46 m	1.03 m	0.72 m	1.0000	76.89%
Sierre, Switzerland	512.45 m	511.10 m	1.68 m	1.12 m	0.9912	78.40%
Death Valley, USA	-85.12 m	-85.24 m	0.78 m	0.45 m	0.8240	89.65%

🌲 Forest and canopy regions

Location	Estimated height	Reference height	RMSE ↓	MAE ↓	Pearson ↑	Accuracy ↑
Western Ghats (Silent Valley), India	410.79 m	401.72 m	5.19 m	4.06 m	0.9959	33.30%
Florida Forest, USA	27.00 m	26.10 m	1.40 m	1.20 m	0.92	65.30%
Black Forest, Germany	1050.00 m	1058.25 m	1.67 m	1.57 m	0.85	70.30%

Global weighted aggregate

The aggregate uses sample-size spatial weighting across all evaluated pixels.

Benchmark	Evaluation extent	RMSE ↓	MAE ↓	Pearson ↑
Global weighted benchmark	12 landscape typologies, more than 10⁷ evaluated pixels	1.83 m	1.32 m	0.94

---

Technology stack

Layer	Technologies	Responsibility
AI and remote-sensing models	PyTorch, GAMUS-tuned Depth Anything V2, semantic segmentation, Modal GPU	Relative/metric height features and land-cover semantics
Backend API and orchestration	C++20, Drogon, CMake, OpenMP, libcurl	Asynchronous ingestion, inference orchestration and scene generation
Geospatial processing	GDAL/OGR + GEOS, Eigen	GeoTIFF I/O, reprojection, affine transforms and raster alignment
Vision and geometry	OpenCV, CGAL, Halide	Masks, morphology, footprint regularization and terrain computation
Building reconstruction	SAT2LoD2, custom height estimation and UV mapping	LoD2 building geometry with source-aligned roof textures
3D packaging	glTF/GLB, Draco	Compact, browser-ready terrain, building and vegetation assets
Frontend	Three.js, Vite, MapLibre GL, 3D Tiles Renderer	Interactive flythrough, inspection and geospatial visualization
Storage	MinIO, AWS SDK for C++ (S3)	Private storage for GLBs and raster exports
Deployment	NGINX, systemd, Ubuntu Linux	Reverse proxy, static hosting and resilient backend services

---

Run the codebase by module

Repository layout

DepthWizard/
├── Elevation_Estimation_Module/
│   └── gis_service/                 # C++20/Drogon backend and reconstruction engine
├── frontend_service/                # Vite + Three.js interactive viewer
├── Model_Training_Scripts/          # Training and evaluation notebooks/scripts
└── README.md

Prerequisites

• Ubuntu 22.04 or 24.04
• GCC/G++ with C++20 support
• CMake, Git and pkg-config
• Node.js 20+ and npm
• MinIO server
• Drogon, Draco, AWS SDK for C++ (s3), and Halide
• GDAL built with GEOS support

Install the distribution-provided dependencies:

sudo apt update
sudo apt install -y \
  build-essential cmake git pkg-config \
  libgdal-dev gdal-bin libgeos-dev \
  libcurl4-openssl-dev libssl-dev zlib1g-dev uuid-dev \
  libjsoncpp-dev libeigen3-dev libopencv-dev libcgal-dev libomp-dev

Drogon, Draco, AWS SDK for C++ and Halide may need to be installed from source or official release packages. See Elevation_Estimation_Module/gis_service/deploy/README.md for the production deployment procedure.

1. Clone the repository

git clone https://github.com/programmer1128/DepthWizard.git
cd DepthWizard

Run the remaining commands from the repository root unless a section says to use a separate terminal.

2. Object-storage module — MinIO

The backend stores generated GLBs and raster exports in a private S3-compatible bucket.

mkdir -p .local/minio-data

export MINIO_ROOT_USER="REPLACE_WITH_LOCAL_ACCESS_KEY"
export MINIO_ROOT_PASSWORD="REPLACE_WITH_LOCAL_SECRET_KEY"

minio server .local/minio-data \
  --address 127.0.0.1:9000 \
  --console-address 127.0.0.1:9001

Keep this terminal running. The local MinIO console is available at http://127.0.0.1:9001.

3. Backend module — elevation and 3D reconstruction

In a second terminal:

cd Elevation_Estimation_Module/gis_service

cp deploy/secrets.env.example deploy/secrets.env

Fill deploy/secrets.env without committing it:

OPENTOPOGRAPHY_API_KEY=REPLACE_IF_USING_REFERENCE_COMPARISON
DEPTHWIZARD_MINIO_ENDPOINT=127.0.0.1:9000
DEPTHWIZARD_MINIO_ACCESS_KEY=REPLACE_WITH_LOCAL_ACCESS_KEY
DEPTHWIZARD_MINIO_SECRET_KEY=REPLACE_WITH_LOCAL_SECRET_KEY

Build and run:

./deploy/scripts/build_backend.sh
cd build-release
./gis_service

The service reads deploy/secrets.env and the validated deploy/depthwizard.env automatically. Test it with:

curl http://127.0.0.1:8081/api/v1/ping

If Halide is installed in a non-standard prefix, configure CMake explicitly:

cmake -S . -B build-release \
  -DCMAKE_BUILD_TYPE=Release \
  -DBUILD_TESTING=OFF \
  -DHalide_DIR=/path/to/halide/lib/cmake/Halide

cmake --build build-release --parallel --target gis_service

4. Frontend module — interactive Three.js viewer

In a third terminal:

cd frontend_service
npm ci
npm run dev

Open http://127.0.0.1:5173. Vite automatically proxies /api requests to the backend at http://127.0.0.1:8081.

To point the development frontend at another backend:

DEPTHWIZARD_BACKEND_URL=http://SERVER_ADDRESS:8081 npm run dev

5. Model training and evaluation module

The production checkpoints are used by the hosted inference services. The training notebooks and the local baseline harness are retained for reproducibility:

cd Model_Training_Scripts
python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install jupyterlab

# Install the versions required by the selected notebook/environment,
# then launch Jupyter or run the baseline harness.
jupyter lab

Available entry points include:

baseline_harness.py
dav2-traina-cls-1.ipynb
dav2-trainb-cls-1.ipynb
dav2-trainc-cls-1.ipynb

6. Run tests

Backend:

cd Elevation_Estimation_Module/gis_service
cmake -S . -B build-tests -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=ON
cmake --build build-tests --parallel
ctest --test-dir build-tests --output-on-failure

Frontend:

cd frontend_service
npm ci
npm test
npm run build

7. Test a GeoTIFF end to end

curl -X POST http://127.0.0.1:8081/api/v1/processor \
  -F "image=@/absolute/path/to/test-image.tif"

The response includes a job UUID and a relative /api/v1/download/... URL for the reconstructed GLB and generated raster products.

---

How it works

flowchart LR
    A[GeoTIFF / PNG / JPG<br/>or map bounding box] --> B[GDAL ingest<br/>validation + reprojection]
    B --> C[Overlapping image tiles]
    C --> D[Modal GPU inference<br/>depth + semantics]
    D --> E[Confidence-aware stitching]
    E --> F{Georeferenced?}
    F -->|Yes| G[Metric calibration<br/>reference terrain fusion]
    F -->|No| H[Relative surface model]
    G --> I[Scene-aware reconstruction]
    H --> I
    I --> J[Terrain mesh]
    I --> K[SAT2LoD2 buildings]
    I --> L[Canopy / trees / shrubs]
    J --> M[UVs + optical textures]
    K --> M
    L --> M
    M --> N[Draco-compressed GLB]
    N --> O[MinIO private storage]
    O --> P[Three.js interactive viewer]

Loading

Supported output behavior

Input	Elevation product	3D behavior
Georeferenced GeoTIFF	Metric DSM and derived raster exports	Metric terrain, LoD2 buildings and vegetation-aware scene
PNG/JPG without spatial metadata	Relative DSM	Relative 3D reconstruction and visualization
Bounding box	Sentinel-2 optical GeoTIFF followed by reconstruction	Best suited to terrain and forest because Sentinel-2 is 10 m/pixel

---

API overview

Method	Endpoint	Purpose
GET	/api/v1/ping	Health check
POST	/api/v1/processor	Process a georeferenced GeoTIFF (image multipart field)
POST	/api/v1/processor/normal-image	Process a non-georeferenced JPG/PNG
POST	/api/v1/global-map/geotiff	Fetch and process imagery for a bounding box
GET	/api/v1/download/{file}	Download generated GLBs and raster products
GET	/api/v1/processor/exports/{uuid}	Inspect raster-export status
POST	/api/height/single	Query height at one location
POST	/api/height/compare	Compare estimated and reference heights

---

Data, attribution and responsible use

• Bounding-box imagery uses Copernicus Sentinel-2 L2A data from the public Earth Search catalogue and AWS open-data bucket.
• Reference elevation may use OpenTopography and other configured DEM providers; their individual terms and attribution requirements apply.
• Bundled vegetation assets and their licences are documented under Elevation_Estimation_Module/gis_service/assets/vegetation/.
• DepthWizard does not require OSM building geometry for reconstruction.
• Validate estimated elevation against authoritative survey data before safety-critical, military, engineering, or emergency-response use.

---

Team

Member	Profile
Aritra Banerjee	programmer1128
Mehul Roy	cubexgani
Avishikta Hajra	avishikta-hajra
Satadru Ghosh	Satadru345
Ishita Mandal	Ishitamandal1595
Sayanti Ghosh	Sayanti6

---

Built for the Smart India Hackathon 2026 · DepthWizard