YOLOX's mAP on swin-L

[LYMDLUT]

As we all know，HTC++ with Swin-G as two-stage detector，used to achieve SOTA on COCOval2017. And DINO with swin-L as Transfomer detector，has achieved SOTA then. Swin-G is much larger than Swin-L, and with resnet50 backbone ,one-stage detector such as YOLOX achieved higher performance than two-stage detector.So it is very important to build the baseline for one-stage detector,to explore with CNN architecture and NMS-based algorithm,can we still achieved the same mAP as DINO?
Whether success or failure,it has great meaning to the next generation of common object detector.

mmYOLO,as the world's leading performace YOLO-series algorithms library, has already supported YOLOX series and Swin series, can you or any other community students, take responsibility for the next generation of object detector?