Gold-YOLO: Efficient Object Detector via Gather-and-Distribute Mechanism: Huawei proposes GD mechanism and MAE pre-training, achieving 39.9% AP and 1030 FPS on COCO
Gold-YOLO proposes the Gather-and-Distribute (GD) mechanism, combining convolution and self-attention operations to enhance multi-scale feature fusion. It is the first to introduce MAE-style unsupervised pre-training in the YOLO series. Gold-YOLO-N achieves 39.9% AP on COCO val2017 and 1030 FPS on T4 GPU, 2.4% AP higher than YOLOv6-3.0-N. Code is open-sourced.
Gold-YOLO improves information fusion in feature pyramid networks through the GD mechanism, using convolution and self-attention for efficient multi-scale feature aggregation and distribution. It is the first to apply MAE pre-training to the YOLO series, enhancing model generalization. It significantly improves accuracy while maintaining real-time performance, providing a new option for industrial deployment.
The core of Gold-YOLO is the Gather-and-Distribute (GD) mechanism, consisting of a Gather stage (multi-scale feature aggregation) and a Distribute stage (feature distribution to each level). Gather uses a mix of convolution and self-attention operations, while Distribute is implemented via lightweight modules. The model adopts MAE-style pre-training, unsupervised learning on ImageNet followed by fine-tuning. Evaluated on COCO val2017, Gold-YOLO-N achieves 39.9% AP at 1030 FPS, and variants like Gold-YOLO-S/L also outperform models of similar scale. A limitation is that only object detection tasks have been validated, not instance segmentation, etc.
Gold-YOLO has direct application value in real-time detection scenarios such as autonomous driving, security surveillance, and industrial quality inspection. Its high accuracy and high frame rate can replace existing YOLO series models, reducing deployment costs. Huawei's open-source code facilitates industrial adoption.
It is recommended to evaluate the replacement cost of Gold-YOLO in existing detection systems, prioritizing scenarios requiring high frame rates (e.g., video stream analysis) and moderate accuracy (e.g., security). Contact the Huawei MindSpore team for optimized deployment solutions.
Future work could explore the transfer of the GD mechanism to other vision tasks (e.g., segmentation, pose estimation) and the benefits of MAE pre-training for larger YOLO models. Actual inference speed and accuracy on edge devices need to be verified.