Built a computer vision system to detect, segment, and reason about objects in grid-based challenge images. The project focused on robust perception plus rule-based spatial decision logic.
- Trained and evaluated Detectron2 instance segmentation models for object-level detection in noisy, tiled image layouts.
- Designed post-processing logic to map detections onto grid coordinates and compute action targets from spatial constraints.
- Iterated on preprocessing, thresholding, and inference-time filtering to improve stability across varied image compositions.
- Implemented a 3-stage pipeline: segmentation, grid-coordinate mapping, and rule-based decision selection.
Code is private for this project, but I can discuss architecture, tradeoffs, and implementation details.