DepthWizard - Single-View Height Estimation and 3D Flythrough
“The challenge is to build software that turns a single 2D satellite picture into a 3D terrain model complete with accurate heights. Users can then fly through this 3D landscape on a web browser to analyze disaster damage, urban structures, and slopes.”
Deconstructed Problem Blueprint
WHAT TO BUILD?
An end-to-end software suite consisting of a monocular depth-estimation AI pipeline, a scale-calibration module using low-resolution DEMs (like SRTM), and an interactive 3D WebGL visualization engine (Three.js/Babylon.js).
WHO IS IT FOR?
ISRO scientists, Department of Space disaster management cells, urban planners, and rescue operations teams.
WHY DOES IT EXIST?
Traditional 3D elevation data requires expensive stereo-imaging, LiDAR, or InSAR radar passes which are costly, scarce, and computationally heavy. Single-view imagery is abundant, but existing AI models only output relative depths without metric real-world height scales.
INPUT → PROCESS → OUTPUT ARCHITECTURE PIPELINE
Single-view optical RGB satellite images (PNG/JPG for relative DSMs or GeoTIFF for absolute DSMs), optional auxiliary low-resolution DEMs (SRTM 30m) or Ground Control Points.
Monocular depth estimation via pre-trained vision models (e.g., Depth Anything, MiDaS), affine scale alignment using auxiliary DEM statistics or coordinate metadata, and displacement-mapped mesh generation.
GeoTIFF absolute Digital Surface Model (DSM), 3D terrain mesh assets, and an interactive browser-based first-person flythrough interface.
CONSTRAINTS
- Must handle domain gaps between natural egocentric training sets and nadir/oblique satellite imagery.
- Must successfully bridge scale-agnostic depth maps into absolute metric elevations using auxiliary sources.
- Rendering pipeline must run smoothly in standard web browsers without crashing under high-poly terrain geometries.
JURY METRICS
- RMSE, MAE, and correlation metrics against reference LiDAR or benchmark data.
- Rendering frame rate (FPS) and visual fidelity during 3D flythrough navigation.
- UI/UX intuitiveness, stability, and ease of standalone deployment.
CORE BOTTLENECK
Accurately converting relative depth predictions (which only know what is closer or farther) into absolute metric heights (meters above sea level) without suffering from severe distortion in flat or heavily sloped terrains.
Why You Should Choose This PS (Pros)
- High national impact for disaster management and ISRO's mapping mandates.
- Clear, highly visual deliverables with a massive 'wow factor' via the 3D flythrough.
- Leverages rapidly advancing open-source monocular depth foundation models.
Challenges & Risks to Watch (Cons)
- Bridge from relative depth to absolute metric scale can be mathematically tricky during live demos.
- Heavy 3D mesh rendering can lag on lower-end hackathon presentation laptops if not optimized.
- Crowded domain if teams rely purely on vanilla open-source depth estimation without remote-sensing domain adaptation.
Recommended Tech Stack & Grand Finale Features
Killer Features to Impress the SIH Jury
View Official SIH 2026 Portal Raw Text▼
Background Accurate Digital Elevation Models (DEMs) and Digital Surface Models (DSMs) are fundamental to urban planning, disaster management, and military reconnaissance. Traditionally, elevation data is acquired through stereo-imaging pairs, LiDAR, or Interferometric Synthetic Aperture Radar (InSAR). These approaches can be cost-prohibitive, dependent on specific sensor availability, and computationally intensive. Single-view height estimation offers an agile alternative, but foundational monocular depth models are trained largely on natural egocentric imagery and predict relative depth. When applied to remote sensing, they face domain gaps, structural variations, and a lack of absolute-scale mapping. Converting relative depth into metric elevation remains a critical challenge, alongside the operational need to transform static elevation profiles into interactive 3D assets that can be navigated in real time.
Description Develop an end-to-end software pipeline that transforms single-view optical RGB remote-sensing images into high-precision elevation maps. The framework must support both non-georeferenced and georeferenced imagery.
The solution should use a pre-trained monocular depth-estimation backbone to generate initial relative-depth maps. For georeferenced imagery, a lower-resolution DEM source such as SRTM or a limited set of Ground Control Points may be used to map scale-agnostic depth features to absolute metric elevations. For non-georeferenced imagery, relative height may be used directly in the visualization stage.
After computing the elevation map, the system should project the original optical image onto a generated 3D terrain mesh and integrate the result with a rendering engine such as Unity, Three.js, or Babylon.js. The interface should support seamless first-person navigation and analysis of structural heights and slopes from arbitrary aerial perspectives.
Key Milestones
Evaluation Criteria:
Expected Solution Deliver a fully integrated software suite with complete source code and technical documentation. The solution must be deployable as a unified module containing the following components: