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
9.29k stars 2.18k forks source link

Predict result compare with yolov4-tiny-3l and yolox-s (Same input data) #1248

Closed er778899789 closed 2 years ago

er778899789 commented 2 years ago

I use same dataset training with yolov4-tiny-3l and yolox-s. Both model mAP50 best score are the same about 95%. But the confidence score in result image, yolov4-tiny-3l is almost better than yolox-s.

But compare with confidence score why yolox-s worse than yolov4-tiny-3l?

Can anyone give me some advice? THX!!

FateScript commented 2 years ago

The score in yolox is not using sqrt function. If socre value is 0.81 in YOLOX, the true score value is 0.9 actually.

er778899789 commented 2 years ago

@FateScript I find confidence score formula in this function.

I print two tensor's value find output[:, 4] get very high score, almost tensor([0.99 ~ 0.98]). And output[:, 5] get tensor([0.94 ~ 0.80]).

Can you explain what the meanings about output[:, 4] and output[:, 5] ?

FateScript commented 2 years ago

centerness and pred_score