Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Amith Kamath Belman

Second Advisor

Fabio Di Troia

Third Advisor

Genya Ishigaki

Keywords

Object Detection, YOLO, Attention Mechanisms, Anchor Assignment, Out-of-Distribution Detection, Traffic Sign Detection

Abstract

In this work, we built a YOLO11-nano architecture fully from scratch without pretrained weights or the Ultralytics framework. We introduced C2DSA (a dual-stream attention module), SAAL (Size-Adaptive Anchor-Assignment Logic), and a contrastive OOD scoring head. C2DSA increased the mAP@0.5 from 0.746 to 0.764 by adding spatial-only self-attention and channel-wise cross-covariance. The OOD head achieves a 0.997 AUROC at very low computational and inference cost. We validated these improvements on the German Traffic Sign Detection Benchmark (GTSDB). These enhancements, combined with test-time augmentation, raised the baseline mAP@0.5 from 0.492 to 0.792 (+61% relative). Crucially, we further validated these enhancements by integrating them with the official pretrained Ultralytics YOLO11-nano through module injection. All of these improvements combined with TTA raised mAP@0.5 from 0.956 to 0.979 (+2.5%).

Download

Available for download on Friday, May 21, 2027

Share

COinS