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
9.4k stars 2.2k forks source link

6 points lower map with yolox-m than yolov5-m, same data augmentation paras on same custom dataset #868

Open jimmyflycv opened 2 years ago

jimmyflycv commented 2 years ago

hello,

I do have tried many pairs of hyp -paras but still with more than 6 points map(0.5) gap here. Do u have any suggestions, since i have viewed all existing issues.

asizk commented 2 years ago

I have the same problem, using a custom data set, about 2 points less on yolox-s compared to yolov5s

Shahji55 commented 2 years ago

I experienced almost the same results. First, I trained on my custom dataset using yolov5m and got mAP of around 62. Then I trained on same custom dataset using yolox_m and got mAP of 56.85.

yoloxm-IOG

sean-wade commented 2 years ago

I experienced almost the same results. First, I trained on my custom dataset using yolov5m and got mAP of around 62. Then I trained on same custom dataset using yolox_m and got mAP of 56.85.

yoloxm-IOG

I get the similar results. Have you found the reason? Thanks~