항공 연료탱크 이물질 탐지를 위한 합성 실제 데이터셋 FOD S2R

FOD-S2R: A FOD Dataset for Sim2Real Transfer Learning based Object Detection Ashish Vashist∗ Qiranul Saadiyean∗ Suresh Sundaram∗ Chandra Sekhar Seelamantula† ∗Department of Aerospace Engineering, Indian Institute of Science (IISc), Bangalore, India †Department of Electrical Engineering, Indian Institute of Science (IISc), Bangalore, India Abstract—Foreign Object Debris (FOD) within aircraft fuel tanks presents critical safety hazards including fuel contami- nation, system malfunctions, and increased maintenance costs. Despite the severity of these risks, there is a notable lack of dedicated datasets for the complex, enclosed environments found inside fuel tanks. To bridge this gap, we present a novel dataset, FOD-S2R, composed of real and synthetic images of the FOD within a simulated aircraft fuel tank. Unlike existing datasets that focus on external or open-air environments, our dataset is the first to systematically evaluate the effectiveness of synthetic data in enhancing the real-world FOD detection performance in confined, closed structures. The real-world subset consists of 3,114 high- resolution HD images captured in a controlled fuel tank replica, while the synthetic subset includes 3,137 images generated using Unreal Engine. The dataset is composed of various Field of views (FOV), object distances, lighting conditions, color, and object size. Prior research has demonstrated that synthetic data can reduce reliance on extensive real-world annotations and improve the generalizability of vision models. Thus, we benchmark several state-of-the-art object detection models and demonstrate that introducing synthetic data improves the detection accuracy and generalization to real-world conditions. These experiments demonstrate the effectiveness of synthetic data in enhancing the model performance and narrowing the Sim2Real gap, providing a valuable foundation for developing automated FOD detection systems for aviation maintenance. I. INTRODUCTION Recent advances in aviation have significantly improved flight safety, operational efficiency, and airport infrastructure. However, as air traffic has increased, safety concerns related to airport environments have become increasingly critical. One of the persistent threats to aviation safety is Foreign Object Debris (FOD), which includes any loose object such as nuts, bolts, tools, or metal fragments that can damage aircraft during takeoff, landing, or ground operations. Small debris such as metal shavings can have severe consequences, including engine failure, fuel leaks, and system malfunctions, ultimately compromising passenger safety and flight integrity. While modern computer vision and AI-based object de- tection techniques have significantly advanced the ability to identify foreign objects in open environments, such as runways and airport aprons, a major challenge remains in detecting FOD in constrained and enclosed areas, such as inside fuel tanks or maintenance bays [11] [29] [62]. These spaces often have poor lighting, complex surfaces, irregular geometries, and limited accessibility, making conventional detection methods less effective. Fig. 1: Unreal Engine editor used to assemble fuel-tank scenes for synthetic FOD data generation, allowing precise control over camera pose, lighting, and object layout. Fig. 2: Blueprint scripting for Dataset generation using Unreal Engine. Furthermore, the development of robust AI models has been hindered by the lack of comprehensive datasets specifically designed for such complex environments. Most existing FOD datasets focus on outdoor conditions and do not capture the visual and spatial constraints unique to internal aircraft com- partments, such as fuel tanks. Collecting and annotating these datasets in aviation settings is labor intensive, expensive, and often restricted by safety regulations. Furthermore, real-world datasets tend to be limited in size and diversity, often under- representing critical FOD classes, which can introduce bias and degrade model performance. To address these limitations, synthetic data generation through high-fidelity simulations has gained traction as a viable alternative. Using platforms such as Unreal Engine, CARLA, and NVIDIA Isaac Sim [12] [10] arXiv:2512.01315v1 [cs.CV] 1 Dec 2025 [31], researchers can create photorealistic environments [28] and automatically generate richly annotated data, including bounding boxes, semantic masks, and depth information. These simulators provide fine-grained control over variables such as lighting, object placement, and camera perspectives, allowing for the simulation of diverse and rare scenarios. However, models trained solely on synthetic data often suffer from a domain gap, in which their performance on real-world data is significantly compromised. To bridge this gap, the Sim2Real transfer-learning approach was adopted by researchers [48]. In this method, the model is first trained on a synthetic dataset, and then fine-tuned on a limited rea

This content is AI-processed based on open access ArXiv data.