Improving object detection accuracy by incorporating attention mechanisms into a DETR model

Abstract

Object detection is a critical task in computer vision, with applications spanning autonomous driving, surveillance, and robotics. The Detection Transformer (DETR) model has recently gained attention for its end-to-end object detection approach using transformers, eliminating the need for traditional components like non-maximum suppression and anchor boxes. However, DETR's accuracy can be limited in complex scenes, particularly when detecting small or overlapping objects. This research explores enhancements to object detection accuracy in DETR by refining its attention mechanisms. In this study, we adjust the multi-head attention configurations in DETR’s transformer layers, varying the number of attention heads in both the encoder and decoder to capture finer inter-object relationships and enhance the model's focus on relevant features. Fine-tuning experiments were conducted on the COCO dataset, with evaluation metrics including mean Average Precision (mAP), Intersection over Union (IoU), and F1 score on a reserved test set. The results demonstrate that our attention-based modifications yield improved performance, with a notable increase in mAP and IoU scores, alongside a 12% improvement in F1 score and a 10% boost in detection accuracy on test data compared to the baseline DETR model. These enhancements were particularly effective in scenes with high object density and occlusion. This work contributes to the field of object detection by illustrating how targeted adjustments to attention mechanisms within transformer-based architectures can yield significant accuracy gains. Our findings highlight the potential of attention mechanism optimization to advance object detection performance, providing a foundation for future research in real-world applications.