YOLOv11 Architecture Optimization

Small objects — distant or partially occluded PPE items, for instance — are notoriously hard for object detectors to catch, and that gap was the focus of my undergraduate (S1) thesis at Universitas Brawijaya: could attention mechanisms improve YOLOv11's small-object performance in industrial safety settings? To find out, I modified the YOLOv11 architecture by inserting a Coordinate Attention module into the neck, letting the network encode positional information alongside channel attention, then trained and benchmarked the modified model against the stock baseline on an industrial PPE dataset.

Coordinate Attention works by factorizing spatial attention into two 1D encodings, so the neck retains long-range positional context cheaply without a heavy compute cost. I ran the experiments in PyTorch, comparing the modified neck against the baseline and visualizing metrics with Pandas and Matplotlib, and the modified architecture demonstrated measurable gains on small PPE object detection. The research became my thesis and contributed to a final GPA of 3.58 / 4.00.

The clearest takeaway was that targeted architectural changes, grounded in the actual failure modes of the baseline, beat blindly scaling the model. Understanding why small objects were being missed in the first place is what told me where attention would genuinely help — and where it would just add parameters.