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
9.41k stars 2.21k forks source link

YOLOX_M isn't better than YOLOX_S and YOLOX_tiny. #854

Open tamir12312 opened 3 years ago

tamir12312 commented 3 years ago

Hi guys, I am training and testing YOLOX on my data. Until now, I used YoloV4 from darknet (Alexa's repository). Somewhy, I can't get the same results with the YOLOX models. I tried tiny_yolo, yolox_s, and yolox_m models. all failed to reach the same f1 score of yolov4. and surprisingly, when I increased the model size (by taking the yolox_m model for example), the results didn't improve at all. any suggestions on what can I do next to get better results? and any insights regarding why yolox_m isn't better than yolox_s?

Joker316701882 commented 3 years ago

@tamir12312 Maybe you need to adjust the data augmentation, like disabling the Mixup. I guess your dataset is not difficult to train, thus using a larger mode will lead to over-fitting. So I suggest you keep using yolox-s and try to adjust the degree of data augmentation.