SegAttnDetec: A Segmentation-Aware Attention-Based Object Detector
Harish Sundaralingam, Tharrengini Suresh, Thangarajah Akilan · Procedia Computer Science · 2025
Object detection (OD) has emerged as a cornerstone of computer vision applications, with deep learning (DL) driving significant advancements. While modern OD algorithms excel in identifying objects, they often falter when confronted with small objects in intricate scenes. To address this challenge, we introduce SegAttnDetec, a novel framework that leverages semantic segmentation to enhance object detection performance. By fusing semantic segmentation-aware features with the backbone of an OD model and incorporating an attention-gating mechanism, SegAttnDetec enables the model to capture richer, more refined features, leading to substantial improvements in small object detection. Notably, our approach achieves a remarkable 28.5% increase in pedestrian hard category detection and a 38.8% improvement in cyclist hard category detection on the KITTI benchmark dataset; hence, reaching an overall mean average precision (mAP) of 83.5% in all the categories of the same dataset.