Performance of YOSO

[canglangzhige]

Thanks for your work!

The paper points that YOLO delivers competitive performance compared to state-of-the-art models.

In fact, currently, much better results may be achieved both for COCO and Cityscapes datasets. For COCO dataset the PQ = 59.5 is obtained by OpenSeeD (see: https://paperswithcode.com/sota/panoptic-segmentation-on-coco-minival) and PQ of about 70 may be achieved for the Cityscapes dataset (https://paperswithcode.com/sota/panoptic-segmentation-on-cityscapes-val).

The presented approach reaches only 46.4 and 52.5, respectively, far from the state-of-the-art solutions.

It seems that YOSO has sacrificed precision heavily in pursuit of real-time performance？