DQ-DETR: Dynamic Query for Tiny Object Detection

Addressing the shortcomings of previous models when detecting tiny objects, DQ-DETR brings an innovative approach that adapts query numbers dynamically, directly influenced by the object scales and imbalanced instance numbers in images. It consists of a categorical counting module, feature enhancement, and an innovative query selection mechanism, which together propel its performance to new heights, as evidenced on the AI-TOD-V2 dataset.

• Innovative Approach to Tiny Object Detection: Adjusting to the unique challenges posed by small-scale objects.
• Dynamic Adjustment of Queries: Modifying the number of queries based on the objects’ sizes and densities.
• High-Performance Metrics: Setting new standards in accuracy for tiny object detection.
• Compatibility with Aerial Datasets: Specially tailored for datasets dominated by tiny objects.

Through DQ-DETR, light is shed on the underlying issues that previously hindered DETR-like systems in detecting minuscule objects. The study presents a practical and effective solution that could dramatically influence practices and outcomes in surveillance, environmental monitoring, and aerial imagery analysis.