OpenCV Maintainer · Senior Computer Vision Engineer
I'm an OpenCV maintainer working on the DNN engine, execution architecture, ONNX operator support, and CPU/CUDA performance. As a Senior Computer Vision Engineer at Bigvision, I also build 3D perception systems for reconstruction, localization, and camera–IMU calibration.
So far that's 140 merged pull requests across the OpenCV organisation and 486 more reviewed.
- Execution architecture: Separated graph nodes from executors and added per-operator backend dispatch and CUDA execution. PR #29286
- Inference performance: Optimized convolution layouts, attention fusion, and LSTM execution. Selected results below.
- Learned features: Added the DISK feature extractor, with regression tests and accompanying test data. PR #29073
| Workload | Before → after | Implementation |
|---|---|---|
| ResNet50 | 14 → 7.6 ms | Convolution block layout |
| BERT | 26.3 → 9.15 ms | Attention fusion and thin GEMM |
| CRNN LSTM | 17.5 → 2.8 ms | LSTM optimization |
CPU measurements on an Intel i9-14900KS using default thread settings. These are results for the linked workloads and configurations; see each PR for details.
OpenCV currently passes 91% of the 1792-test ONNX conformance suite (1629 of 1792),
up from 37% in September 2025, when I imported the complete suite. Subsequent work includes If, Loop, and Scan operator support.
- Multi-room reconstruction and localization: Built a system achieving under 1° rotation error and 1.5 cm position error across a full home, registering room sub-maps into a metric world frame and evaluating against ArUco ground truth.
- Visual-inertial tracking: Fuse monocular video and IMU data, including camera–IMU extrinsics and time synchronization.
- Public projects: hazard-detector · spatial_query_engine
I also mentor OpenCV GSoC projects in SLAM, 3D visualization, and learned feature matching.
calibsense translates camera-calibration uncertainty into measurement error in millimetres. It diagnoses poorly constrained parameters and flags cases where the available evidence cannot support a reliable forecast.
- VGGT vs. VGGT-Ω: A Complete Guide to Feed-Forward 3D Reconstruction — LearnOpenCV, August 2026.
- Accelerating OpenCV on Graviton: the COOL Framework (archived) — AWS Physical AI Blog, April 2026.