Real-time Object Detection Based on Mamba and YOLOv8
Jie Wang, Wenyi Zhao, Caixi Liu, Huihua Yang, Wei Hong Xu · 2024
Amidst the swift progress in deep learning, the YOLO series has redefined the standards for real-time object detection. The incorporation of the Transformer architecture into YOLO models has significantly broadened their receptive fields and enhanced their feature extraction capabilities, thereby markedly improving detection performance. Nonetheless, this enhancement also leads to increased computational demands. To address this, our manuscript proposes an improved YOLOv8 object detection model that integrates a State Space Model (SSM) to overcome its previous constraints in accuracy and speed. Specifically, we've proposed the MambaCSP module and embedded it into the neck and decoupled modules. By employing depthwise separable convolutions and channel shuffle operations, we've tailored branches for both classification and regression tasks. These innovations not only bolster the model's feature extraction prowess but also maintain its computational efficiency. Our proposed method has demonstrated an average improvement of 1 % on the COCO dataset compared to the YOLOv8 model.