Inverted Bottleneck Convolution Module for Yolov8

Object detection remains a pivotal challenge in computer vision, with the YOLO series establishing itself as the industry standard for efficient solutions. Concurrently, the inverted bottleneck convolution module has been instrumental in the success of the MobileNet family, enabling efficient feature extraction. In this research, we integrate the inverted bottleneck convolution module into the state-of-the-art YOLOv8 object detection model, strategically incorporating it into the low-resolution regions of the network. The proposed model achieves superior performance on the COCO dataset, outperforming the original YOLOv8 by 1% in $\mathrm{m}\mathrm{A}\mathrm{P}^{50:95}$ score while employing fewer parameters and requiring fewer Giga FloatingPoint Operations per Second (GFLOPS), contributing to a more lightweight and efficient architecture. Remarkably, the proposed model exhibits comparable inference speeds on various hardware platforms, ensuring real-time performance and practical deployment capabilities. This balance between accuracy and efficiency positions our model as a compelling alternative to YOLOv8, offering superior object detection performance while maintaining resource-efficient operations. Our work highlights the potential for cross-pollination of architectural innovations across different families of convolutional neural networks, paving the way for further exploration and optimization of efficient object detection models.