Algorithm & Software Engineer | Computer Vision & 3D Depth | Evaluation & Benchmarking

Mask2PolyMin turns noisy segmentation masks into clean polygons with a minimal number of segments and sub-pixel corners.
Benchmarked against OpenCV's approxPolyDP: comparable IoU and Hausdorff distance, without corner-cutting bias.
pip install mask2polymin
If you find it useful, a star helps others discover it!

cv::convexHull spent most of its time in std::sort (O(n log n)), even though the hull algorithm itself is linear. A mentee from Kamatech replaced it with a bucket sort; I extended that into a dispatcher that picks the fastest method for each input:

• Counting sort when the coordinate range is small relative to the point count
• Radix sort (stable, 8-bit digits) when there are enough points per pass
• std::sort otherwise

Issue #28846
PR #28868 (mentee's bucket sort; fixes and improvements by me)
PR #30196 (my follow-up: dispatcher, radix sort, float support)
Both PRs are under review.

How far can you trust camera-based depth? I compared stereo (OpenCV SGBM) and monocular depth (Depth Anything V2) against LiDAR ground truth on KITTI drives. Built during the 3D Ascension program by Think Autonomous.

• bev_from_depth: bird's-eye occupancy maps from each source, scored by IoU per distance band. Stereo reaches 0.59 IoU at 0–20 m but drops to 0.23 at 40–70 m; mono stays below 0.25.
• mono_vs_stereo_vs_lidar: per-frame RMSE and nearest distance per detected person or car, compared with LiDAR.

• 📊 Expertise in computer vision, 3D/depth, computational geometry, combinatorics, graph and route planning
• 💡 Worked on Excel Copilot at Microsoft (2024–25)
• 🏆 Winner of RWTH Aachen Programming Competition for Students (sponsored by Sun Microsystems)

• Languages: Python, C, C++, C#, TypeScript
• Domains: Computer Vision, 3D/Stereo Depth, Computational Geometry, Graph Algorithms, Maps
• Frameworks & Tools: OpenCV, NumPy, PyTorch, Depth Anything, YOLO, pytest, Docker, Linux, Git

• 💼 LinkedIn