Finding Tiny Objects More Reliably in Crowded Aerial Images

What is it about?

This study focuses on detecting very small objects in high-resolution aerial images, such as vehicles, ships, airplanes, and other tiny targets that occupy only a few pixels. These objects are difficult to detect because their visual features are weak, easily confused with background clutter, and often densely distributed. We propose FDQ-Det, a detection framework that combines three complementary ideas. First, it estimates an object density map to understand where objects are likely to appear and how densely they are distributed. Second, it uses this density information to suppress background-dominated high-frequency noise while preserving object-related structures. Third, it regularizes the object queries used by Transformer-based detectors so that they are less sensitive to imperfect density guidance. Together, these components help the detector better distinguish tiny objects from complex backgrounds and reduce missed detections in dense aerial scenes.

Featured Image

Photo by Max Böttinger on Unsplash

Photo by Max Böttinger on Unsplash

Why is it important?

Tiny object detection is important for many aerial and remote sensing applications, including traffic monitoring, disaster response, environmental observation, and wide-area surveillance. In these scenarios, missing small targets can significantly reduce the reliability of an automated system. Existing detectors often struggle because tiny objects have limited visual information and are easily overwhelmed by background noise. FDQ-Det addresses this problem by explicitly modeling object density and using it to guide both feature processing and query initialization. This allows the model to focus more effectively on object-support regions while reducing interference from cluttered backgrounds. Experiments on AI-TODv2 and VisDrone show that FDQ-Det improves detection accuracy, especially for very tiny and tiny objects, and provides better object coverage in dense scenes. The method also operates only during training for its adversarial regularization component, so it improves robustness without adding extra inference-time cost.

Perspectives