Strange predictions for YOLOv4-tiny

[karinaodm]

Hellow, thank you for the wonderful work.

I converted YOLOv4-tiny model ( weights cfg ) using sh ./prepare_cfg.sh yolov4-tiny.cfg yolov4-tiny_temp.cfg python ./convert_v4.py -n coco.names -c yolov4-tiny_temp.cfg -w yolov4-tiny.weights -m yolov4.mlpackage -l RGB

During model inference several boxes for each object are obtained. Video example https://drive.google.com/file/d/1TdCv0Ma6Vj7BWCiI7-1Z-BvzAcQrk05G/view?usp=sharing

However, if I use tensorflow-yolov4 lib for postprocessing of coreml model output I get only one box per object. I think, there is an error in computing of bounding boxes for YOLOV4-tiny.

https://github.com/PROGrand/yolo-coreml/blob/f96b07b2357c1881239b692a792e1533c2d86587/ios/yolov4/YOLO.swift#L174 and https://github.com/hhk7734/tensorflow-yolov4/blob/4f49523210325d8e68bd9ff2671ca003ae80d989/c_src/py_box.cpp#L417 give different result for the same inputs.