Transformer-based End-to-End Object Detection in Aerial Images

Transformer models have achieved significant mile-stones in the field of Artificial Intelligence in recent years, primarily focusing on text processing and natural language processing. However, the application of these models in the domain of image processing, particularly on aerial images data, is actively research. This study concentrates on the experimental evaluation of Transformer-based models such as DETR, DAB-DETR, and DINO on the challenging Visdrone dataset, which is also essential for aerial image data processing. The experimental results indicate that Transformer-based models exhibit substantial potential, especially in object detection on aerial image data. Nevertheless, their application is not without challenges, including low resolution, dense object occurrences, and environmental noise. This work provides an initial glimpse into both the capabilities and limitations of Transformer-based approaches within this domain, with the aim of stimulating further development and optimization for practical applications, including traffic monitoring, environmental protection, and various other domains.