Senior Computer Vision Engineer · OpenCV Maintainer
I build 3D perception systems and make inference engines faster. My work spans SLAM, reconstruction, camera–IMU calibration, ONNX operator coverage, and CPU/CUDA optimization.
- OpenCV: Maintainer working on the DNN engine, graph architecture, ONNX coverage, and CPU/CUDA performance.
- 3D perception: Built a multi-room reconstruction and localization system with under 1° rotation error and 1.5 cm position error across a full home.
- Inference: Improved ResNet50 from 14 ms to 7.6 ms, BERT from 26.3 ms to 9.15 ms, and CRNN LSTM from 17.5 ms to 2.8 ms.
- ONNX: Helped raise OpenCV's conformance pass rate from 37% to 87%.
- Calibration: Created calibsense, which translates calibration uncertainty into task-space error in millimetres.
Across both areas, I work the same way: measure the result, publish it, and state clearly when the evidence does not support a claim.
So far that's 137 merged pull requests across the OpenCV organisation and 258 more reviewed.
Indoor reconstruction and localization. I architected a room-by-room system that registers independently reconstructed sub-maps into one metric world frame by solving inter-room SE(3) extrinsics. ArUco ground truth and multi-view fusion keep drift measurable at every merge step.
Visual-inertial tracking. I fuse monocular video with IMU streams for scale-consistent, gravity-aligned tracking, including camera-to-IMU extrinsics and time synchronization. I also use VGGT and Pi-Long in low-texture scenes where feature-based tracking fails.
Sensor and calibration feasibility. Before implementation, I evaluate whether a proposed camera and IMU configuration can meet the required accuracy, then validate its calibration against reprojection error and fiducial ground truth.
Video to simulation. I turn road footage into metric, navigable simulations using recovered camera geometry, 3D Gaussian splats, collision meshes, physics, and generated assets.
Hands-on stack: ORB-SLAM3, COLMAP, MASt3R, VGGT, Pi-Long, LightGlue, Open3D, ElasticFusion, ACE Zero, Depth Pro, and Kalibr—built, modified, and evaluated on EuRoC and proprietary captures.
Source · PyPI · Documentation
A reprojection error does not tell you whether a system can measure a part to half a millimetre. calibsense retains the full factored parameter covariance, identifies unconstrained directions through null-space analysis, and propagates uncertainty into task space. A 100 mm feature at 800 mm, for example, may read 0.199 mm expected error and ±0.485 mm at 95% confidence. When the evidence is insufficient, it reports that no defensible forecast can be made.
It supports hand-eye calibration, ChArUco, checkerboard and circle-grid input, and JSON, PDF, and self-contained HTML reports.
- Added the DISK learned feature extractor, including regression tests and test data.
- Review and merge work across point clouds, SLAM, visual odometry, loop closure, bundle adjustment, surface reconstruction, and 3D visualization.
- GSoC 2026 mentor for Modular SLAM, a lightweight 3D viewer, and a learned feature extraction and LightGlue matching pipeline.
- #29286: Split graph nodes from executors, added per-operator backend dispatch, and brought CUDA into the new engine with on-device residency. ResNet50 runs end to end at 1.17 ms FP16.
- #29341: Removed the legacy DNN engine and made the new graph engine the default.
- #30081: Added a unified N-dimensional data-movement engine in
core. - Also added custom-layer support and ONNX Runtime as an optional execution engine.
| Work | Result |
|---|---|
| Convolution block layout | ResNet50: 14 → 7.6 ms |
| INT8 block layout | ResNet50-QDQ: 11.5 → 6.6 ms |
| Attention fusion and thin GEMM | BERT: 26.3 → 9.15 ms |
| FlashAttention graph fusion | OWL-v2: 1078 ms, versus ONNX Runtime at 1411 ms |
| LSTM optimization | CRNN: 17.5 → 2.8 ms |
reserveKVCache() |
Qwen2.5-0.5B: 4.25 → 21.03 tok/s at 512 tokens |
CPU results were measured on an i9-14900KS with threads pinned to P-cores. Each result is reproducible from its linked pull request.
OpenCV currently passes 87% of the 1792-test ONNX conformance suite (1552 of 1792),
up from 37% in September 2025, when I imported the complete suite and exposed the cases the earlier test set had skipped.
The work included control flow (If, Loop, Scan), attention, and low-bit numeric formats.
- VGGT vs VGGT-Ω: A Complete Guide to Feed-Forward 3D Reconstruction — LearnOpenCV, August 2026. Both models evaluated on the same 4K sequence and RTX 5090.
- Accelerating OpenCV on Graviton: the COOL Framework — AWS Physical AI Blog, April 2026.
- Four OpenCV webinars and the OpenCV 5 launch preview.
I publish negative results as well as wins: failed optimizations, benchmark regressions, and cases where the available data cannot support a reliable claim.