Work · 2024 – 25
Computer Vision R&D Engineer (Junior)
Deed Asia Development Group Ltd · Shiraz, Iran (remote, then on-site)
Oct 2024 – Mar 2025
R&D project for Aerialytic, a Canadian AI solar-design company — monocular depth estimation and 3D roof reconstruction from aerial imagery.
- Built the monocular depth pipeline for aerial roof imagery — benchmarked Depth Anything V1/V2 and Depth Pro, selected Depth Anything V2 Large, and delivered a CLI that tiles 74,880×15,200 px orthophotos into overlapping windows with pyramid-weighted blending for seamless depth maps.
- Designed roof-face segmentation and obstacle removal — polygonised labelled edges into faces (Shapely), median-filled faces to erase chimneys and trees, then fitted per-face 3D planes with RANSAC and weighted SVD (Open3D) for 3D roof reconstruction.
- Rewrote the megatile preprocessing in OpenCV/NumPy with bounding-box-local masks, cutting the median-fill pass from 90 min to 8 s and the full megatile run from 16 h to 5 min.
- Aligned hand annotations to YOLO keypoint line detections with a weighted cost matrix and the Hungarian algorithm; developed 5-class roof-edge classification rules (ridge, hip, valley, rake, eave) reaching an estimated 70–80% accuracy.
- Wrote 39 twice-weekly progress reports; onboarded two new team members and handed off one workstream.