Object detection and instance segmentation are two core tasks in computer vision, playing a crucial role in various applications. The YOLO series models are known for their efficiency and high accuracy. YOLOv7-seg, in particular, demonstrates extraordinary detection and segmentation performance while maintaining high speed. However, after in-depth study, we found two primary problems of YOLOv7-seg while handling segmentation tasks: 1. When the bounding box (bbox) containing more than one objects of the same class, the original segmentation loss of YOLOv7-seg may suppress the activation of other objects of the same class; 2. The origin class-agnostic segmentation mask is easy to confuse among masks of different classes. To address these problems, we propose YOLO-CM. This model effectively resolves the shortcomings of the YOLOv7-seg in segmentation tasks by combining masks of the same class within a bbox during training and outputting class-aware segmentation mask. On the COCO test dataset, our approach achieves significant performance improvements, with a bounding box average precision (APbox) of 0.516 and a segmentation mask average precision (APmask) of 0.416. Source code is available here: https://github.com/Scorbinwen/Yolo-CM .