search

Megvii-BaseDetection / YOLOX

YOLOX is a high-performance anchor-free YOLO, exceeding yolov3~v5 with MegEngine, ONNX, TensorRT, ncnn, and OpenVINO supported. Documentation: https://yolox.readthedocs.io/
Apache License 2.0
9.47k stars 2.21k forks source link

anchor free 疑问 #1413

Open violet2422 opened 2 years ago

violet2422 commented 2 years ago

您好，最近在深深入研究这项工作，也很感谢您的巨大贡献，请问yolox这种anchor free机制在loss计算的时候如何确保浅层特征预测小目标而深层特征预测大目标的呢？

FateScript commented 2 years ago

It's simliar to FCOS, you could check get_assignments in yolo_head.py for more details.

cswwp commented 2 years ago

You can just consider each grid as an "anchor" whose aspect ratio is 1 and area is downsample size. About label assignment, use get_assignments

Joker316701882 commented 2 years ago

@violet2422 SimOTA会根据loss的匹配程度，对目标自动分层。关于这部分的分析，可以参考LLA