Accuracy on COCO in comparison to official (Darknet) YOLOv4 implementation

[Vottivott]

Hi, do you know if the YOLOv4 model in this repository matches the accuracy of the official Darknet implementation of YOLOv4?

The Darknet implementation has 62.8 mAP50 for 416x416 images while for example the tensorflow-yolov4-tflite implementation only has 57.3 when loaded with the same weights file. Do you have any indication whether your implementation has a closer accuracy to the original?

[hhk7734]

Not yet. Are the evaluation formulas and dataset for darknet and tflite the same?

https://github.com/hunglc007/tensorflow-yolov4-tflite/issues/165