Huge difference to regular darknet output

[ItsMeTheBee]

Hey there!

After compiling my custom tiny yolo v3 network for the edgetpu I´m able to run my network but the output is insanely different.
On my custom network there is only one class, anchors and input size remain the same so i don´t know where these issues might be coming from.

Edge Tpu output:

Desired output

Do you have any why this is happening and how to fix this?

[daynauth]

I've been having the same issue too. I've seen a notable loss in accuracy with the conversion from darknet to keras and this may have been propagated onto the edge tpu model.

[ownbee]

I am experiencing the same kind of issue (worse output after conversion). I don't know if it's the inference.py script or the converted model. I've also noticed that the boxes sometimes have negative coordinates.

[Rariusz]

I have the same problem, any solution?

[YsYusaito]

Hi @ItsMeTheBee @Rariusz
I corrected mistakes of utils.py.(Issue #19)
I hope it helps you solve your problem.

Plus, please confirm that 「anchor.txt」 is loaded properly. (I think your bounding boxes are too small)

[stan-guer]

Maybe it has something to do with the missing representative dataset for int8 calibration? This repo for some reason just generates random data.
Compare:
https://github.com/guichristmann/edge-tpu-tiny-yolo/blob/master/keras_to_tflite_quant.py#L21
to
https://github.com/hunglc007/tensorflow-yolov4-tflite/blob/master/convert_tflite.py#L17

Unfortunately, I couldn't get this repository to work yet, so I can't confirm if this is a fix.