Too many false positive bounding bboxes in my YOLOX detection result

你好，我在训练yoloxs的时候，加载自己的数据，结果发现误检框非常多。因为同样的数据当时跑yolov5的时候误检是没有那么多的，所以我感觉应该不是数据问题。然后我就看了下代码在simOTA做assign gt的时候完全没有对负样本进行采样，同时看到这个issue（https://github.com/Megvii-BaseDetection/YOLOX/issues/725）里也提到了这点。那我理解即使我增加background的图片也无法降低误检，因为并没有采样negative anchor point并进行抑制，请问下我的理解是否正确，同时这个问题是否有解决的计划，谢谢。

Hello, I train yoloxs on my own dataset with too many false positive bounding bbox in result. I suppose that it has no problem with the dataset, because much fewer false positive boxes in yolov5s training with the same dataset. I notice that YOLOX does not sample negative anchor point in simOTA. And this issue mention this problem as well(https://github.com/Megvii-BaseDetection/YOLOX/issues/725). So is that means even if the dataset contains background images, YOLOX can not select negative anchor point for improve the false positive problem. Is that correct? Many Thanks for your Apply.

Start by reading the simOTA ground-truth assignment logic and the linked issue #725, then compare how YOLOX handles background or negative locations with the reported YOLOv5 behavior. Done would require a maintainer-confirmed explanation of whether negative sampling is involved and whether a code change is planned.