Computer vision research
Detecting Bézier curves where polylines were not enough.
I built a direct Bézier control-point detection pipeline for noisy maritime imagery, combining curvature-aware RANSAC, non-linear refinement, occlusion recovery, and a new 5,000-image benchmark.
Technology
The required geometry could not be approximated away
The research use case required direct parametric Bézier control points, not a polyline approximation. Maritime imagery added low contrast, partial occlusion, noise, and changing illumination.
No public benchmark measured the required control-point accuracy, so algorithm development and evaluation infrastructure had to advance together.
Built the detector and its benchmark
I owned the geometric detection pipeline, curvature-aware sampling, Levenberg-Marquardt refinement, occlusion recovery, benchmark generation, and Python and ROS integrations.
The work focused on measuring the geometric property directly rather than relying on gradient orientation as an unreliable proxy.
Architecture
Evaluation and detection advanced together
The algorithm, controlled data generation, geometric metrics, and integration surfaces were designed as one research system because no existing benchmark measured the required control-point accuracy.
- 01
Constraint
Orientation-weighted RANSAC samples became biased and required too many iterations in noisy scenes.
Decision
Reject degenerate minimal sets using a curvature prior derived from the circumscribed arc.
Impact
The benchmark required approximately 60% fewer iterations for 99% detection probability.
- 02
Constraint
Occlusion produced gaps that caused otherwise correct curve candidates to lose too many inliers.
Decision
Project a refined Bézier through long gaps and retest pixels along the expected path against the edge map.
Impact
Correct detections were recovered on approximately 15% of medium-difficulty benchmark images.
- 03
Constraint
Existing line and wireframe datasets could not measure control-point accuracy.
Decision
Build a calibrated synthetic benchmark with known control points, controlled noise, occlusion, and contrast variation.
Impact
More than 5,000 annotated images supported repeatable mAP and Hausdorff-distance evaluation.
Outcome
Sub-pixel geometry with a benchmark others could reuse
- The detector tolerated 40% outliers at the medium benchmark tier.
- Non-linear refinement delivered sub-pixel control-point estimates.
- The implementation exposed C++, Python, REST, and ROS integration paths.
- The benchmark was reused by three downstream research pipelines.
The benchmark became as important as the detector. When no existing metric captures the required geometry, building credible evaluation is part of building the algorithm.
Abrar
Start a project
Need a computer-vision system grounded in measurable geometry?
I can help build the algorithm, benchmark, integration surfaces, and evidence needed to move beyond a promising prototype.
Discuss your vision system