Improved object detection via large kernel attention

Abstract

Accurate and real-time object detection is crucial for autonomous vehicles. For in-vehicle computing platforms, a giant model is difficult to achieve in real time and also increases the hardware cost. Moreover, a lightweight model built from a lot of depth-wise separable convolution layers cannot achieve the accuracy required for autonomous driving. We introduce Large Kernel Attention (LKA) technology to decouple the large kernel convolutions. It can combine high accuracy with small computational cost. Furthermore, we use LKA as the basis for designing a new module (Res-VAN) that can be used to build backbone networks. This study focuses on how the Res-VAN module can be deployed to improve the YOLOv5 in order to reduce the computational effort of the model, but maintain its accuracy. We named the model “LKA-YOLO” and validated it on PASCAL VOC dataset, MS COCO 2017 dataset and VirDrone2019 datasets. Experimental results show that LKA-YOLO reduces the computation of YOLOv5 by 57.5% (16 GFLOPs vs. 6.8 GFLOPs). And the feature extraction ability of LKA-YOLO is stronger than YOLOv5 on small datasets. The results on VirDrone2019 dataset show that our proposed method has a great advantage in dealing with tiny objects. Meanwhile, comparing with YOLO family on the COCO 2017 dataset, the metrics of LKA-YOLO are also outstanding.