Bilkent University
Department of Computer Engineering
CS 590/690 SEMINAR
Spatial Contextual Object Detection via Graph Neural Networks
Sinan Çavdar
Ph.D. Student
(Supervisor:Prof.Dr.Selim Aksoy)
Computer Engineering Department
Bilkent University
Abstract: Object detection in aerial images remains challenging due to substantial variations in object scale and orientation, cluttered backgrounds, and limited visual information for small objects. Spatial relationships among neighboring objects provide contextual information that can help improve object detection. We propose a framework for incorporating explicit spatial context into oriented object detection, using Oriented R-CNN as the baseline. The proposed framework combines an enhanced feature pyramid network (FPN), a graph neural network over selected region proposals, and a contextual loss that supervises relative object geometry. The FPN enhancement aims to preserve the fine spatial details needed to detect small objects while retaining high- level semantic information. Initial experiments evaluate this component independently, using the DOTA 1.5 validation split as the test set. Modifying only the FPN increases mean average precision (mAP) from 72.12% to 72.49% across all 16 classes, an improvement of 0.37 percentage points. Excluding the container-crane class, mAP increases from 76.32% to 77.17% across the remaining 15 classes, an improvement of 0.85 percentage points. The contextual loss has been evaluated in preliminary simulations using synthetic proposals, while ongoing work focuses on integrating this loss and the graph neural network into the detector.
DATE: November 09, Monday @ 15:00 Place: EA 502