Project
3D Photogrammetry Pipeline — DRDO CAIR
Complete Structure-from-Motion pipeline for city-scale 3D reconstruction from drone imagery. Built during a DRDO CAIR internship under Dr. Sumit Veerwal — from manual SIFT implementation to Dockerised industrial pipeline.
Problem Statement
Reconstruct geometrically accurate, metric-scale 3D models from aerial drone footage — targeting city-scale digital twin for terrain mapping and reconnaissance applications.
System Design
SIFT implemented from scratch (DoG pyramid → scale-space extrema → orientation assignment → 128-D descriptors), then industrial pipeline: COLMAP feature extraction → exhaustive matching → incremental SfM → dense reconstruction → Meshlab metric scaling. Entire toolchain Dockerised (~3.5GB) with CUDA support.
Key Challenges
Resolving GPU passthrough and GUI forwarding in Docker containers, handling low-texture featureless regions, scaling models without ground-truth references, and processing city-scale datasets within memory constraints.
Results
5 multi-scale datasets processed (25–274 images/frames). City-scale reconstruction from 83-frame aerial video demonstrated. Dockerised pipeline tested on DRDO lab computers. Internship at DRDO CAIR, Bengaluru (June–July 2025).