YOLO-MS: Rethinking Multi-Scale Representation Learning for Real-Time Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Real-time object detection is crucial for many applications.
• Existing models face challenges in effectively representing multi-scale features.

Purpose of the Study:

• To develop an efficient and performant object detector, YOLO-MS.
• To enhance multi-scale feature representation in real-time object detectors.

Main Methods:

• Investigated the impact of multi-branch features and varying kernel sizes on object detection performance.
• Developed a novel strategy for enhancing multi-scale feature representations.
• Trained YOLO-MS on the MS COCO dataset from scratch.

Main Results:

• YOLO-MS outperforms state-of-the-art real-time object detectors including YOLO-v7, RTMDet, and YOLO-v8.
• YOLO-MS-XS achieves over 42% AP on MS COCO, surpassing RTMDet by 2% with similar model size.
• The YOLO-MS module improves YOLOv8-N's APs, APl, and AP by significant margins with reduced parameters and MACs.

Conclusions:

• YOLO-MS offers a significant advancement in real-time object detection.
• The proposed method effectively enhances multi-scale feature representation.
• YOLO-MS can be integrated as a plug-and-play module to boost performance of other YOLO models.