Gold-YOLO: Efficient Object Detector via Gather-and-Distribute Mechanism

In the past years, YOLO-series models have emerged as the leading approachesin the area of real-time object detection. Many studies pushed up the baselineto a higher level by modifying the architecture, augmenting data and designingnew losses. However, we find previous models still suffer from information fusionproblem, although Feature Pyramid Network (FPN) and Path Aggregation Network(PANet) have alleviated this. Therefore, this study provides an advanced Gatherand-Distribute mechanism (GD) mechanism, which is realized with convolution andself-attention operations. This new designed model named as Gold-YOLO, whichboosts the multi-scale feature fusion capabilities and achieves an ideal balancebetween latency and accuracy across all model scales. Additionally, we implementMAE-style pretraining in the YOLO-series for the first time, allowing YOLO-seriesmodels could be to benefit from unsupervised pretraining. Gold-YOLO-N attains anoutstanding 39.9% AP on the COCO val2017 datasets and 1030 FPS on a T4 GPU,which outperforms the previous SOTA model YOLOv6-3.0-N with similar FPS by+2.4%. The PyTorch code is available at https://github.com/huawei-noah/EfficientComputing/tree/master/Detection/Gold-YOLO, and the MindSpore code is available at https://gitee.com/mindspore/models/tree/master/research/cv/Gold_YOLO. © 2023 Neural information processing systems foundation. All rights reserved.