DG-YOLOT: A LIGHTWEIGHT DENSITY GUIDED YOLO-TRANSFORMER FOR REMOTE SENSING OBJECT DETECTION

Deep learning-based object detection methods in natural image datasets have demonstrated remarkable accuracy and lower error rates than those of humans. As a result, they have gained significant attention in the field of remote sensing imagery. However, direct transferablity of these methods in remote sensing images face challenges such as scale variations, complex object distributions, and arbitrary orientations. In order to address these challenges, we propose the transformer-based object detector named as DG-YOLOT where we use a guided self-attention mechanism with YOLOv5 to enhance the potentiality of training with minimum computation. Instead of employing uniform size patches like the conventional vision transformer, we leverage density map patches which facilitates the extraction of diverse contextual information related to objects within the image, enhancing the differentiation capability of our model. Through extensive experiments conducted on the DOTAv2.0 dataset, our proposed model has demonstrated superior performance with 57.91% mean Average Precision (mAP) compared to other state-of-the-art object detectors.