Yet-another-efficientdet-pytorch: WORSE THAN TENSORFLOW OFFICIAL RESULTS INFERRING CUSTOM DATASETS

Created on 6 May 2020 · 3Comments · Source: zylo117/Yet-Another-EfficientDet-Pytorch

https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch/issues/258#issue-612721607
It's the truth 'YET' version is much worse than tf official version... with the same output settings..

Source

qtw1998

Most helpful comment

Actually the tf version has so so so so many tricks for data augmentation.....
Yet version has nothing..

It's like playing with axes at the door of Luban home..

qtw1998 on 6 May 2020

👍3

All 3 comments

Even I set the settings as iou_threshold=0.9, threshould=0.1, --force_input_size=20000(huhu)
the areas have been detected be much much worse than official
YET VERSION -- iou_threshold=0.9, --threshould=0.1, --force_input_size=20000
output
you could see lotta traffic signs and lights have been thought as background
TF OFFICIAL VERSION --iou_threshold=0.5, --threshould=0.5, --force_input_size=1920(origin input size)
68747470733a2f2f736974652d70696374757265732e6f73732d65752d776573742d312e616c6979756e63732e636f6d2f6c6d6170332e6a7067
Well... I feel it's hard to trust the YET VERSION...

qtw1998 on 6 May 2020

Actually the tf version has so so so so many tricks for data augmentation.....
Yet version has nothing..

It's like playing with axes at the door of Luban home..

qtw1998 on 6 May 2020

👍3

is data augmentation the only difference between this repo and tensorflow repo ?