Hierarchical Fine-Grained Aerial Object Detection

A research team has introduced ExpertDet, a detection scheme designed to distinguish visually similar aircraft and ship models in aerial imagery by incorporating structured expert knowledge into the detection pipeline [1]. The work addresses a persistent limitation in remote sensing: standard object detectors, trained with single-label supervision, often fail to separate model-level categories that share broad class characteristics but differ in subtle structural details [1][2]. For a specific airframe such as a Boeing 787, attributes like wing shape and hierarchical relationships — for instance, plane to wide-body airliner — carry discriminative information that coarse-grained methods overlook [3][4]. ExpertDet operationalizes this insight through two components. Vision-aware Masked Attribute Modeling, or VMAM, reconstructs randomly masked attributes from visual features, forcing the model to attend to fine structural distinctions [2][3]. Hierarchical Visual Instance Promotion, called HierVIP, constructs a visual prototype tree from taxonomy relations and applies constraints that preserve semantic continuity across hierarchy levels while sharpening inter-class boundaries [2][4]. To support evaluation, the authors curated the PSP benchmark, which stands for Precise recognition of model-specific Ships and Planes [1]. The dataset contains 106 ship classes and 30 airplane models, making it the largest collection of model-specific categories among existing aerial object detection datasets [1][2]. The benchmark was used to test several state-of-the-art algorithms, and ExpertDet consistently outperformed other fine-grained competitors across multiple hierarchy levels [1][4]. The challenge of fine-grained aerial detection has drawn wider attention in the computer vision community. A separate 2024 study introduced Orthogonal Mapping, a method that mitigates semantic confusion by assigning orthogonal category prototypes to each class and background, achieving a 4.08 percent improvement in mean Average Precision over the FCOS detector on the ShipRSImageNet dataset [5]. That approach required only a single line of code modification and was validated on three fine-grained datasets: FAIR1M, MAR20, and ShipRSImageNet [5]. The ExpertDet authors have released the PSP dataset, benchmark, and code on a public project page [1][2].