bug ghost detections while using yolov4-tiny-relu

[fogelton]

Hi, I encountered one bug while using the detector over video. I detects 'ghosts'. I noticed that it draw bounding boxes where people are not present. maybe it is something with the conversion to bounding boxes and it puts another for larger objects on wrong locations... or something with anchors.
It can be clearly see that the ghost detection follow large objects lower within the image.

Probably you can find out from the data I attach
debugging.zip

You can see 3 ghosts detections on the images in zip file, I also attach the output of the network, the candidates before and after postprocessing and also the predicted bounding boxes as pickle
so you can find the bug
if you need any further assistance from me, it will be my pleasure

ps: sorry for not doing the pull request I was on holiday

[hhk7734]

What is your yolov4 version?

[fogelton]

I do not know.
how to find out?
these are the weights I downloaded somewhere, maybe from you...

yolov4-tiny-relu.zip

if you share the latest weights with me, I can test it again.
I use the -a flag and I have two subgraphs getting about 60ms per image on edgetpu

[hhk7734]

python3 -m pip show yolov4

[fogelton]

i was using 2.0.0 after updating to 2.0.2, the ghosts are still there

[hhk7734]

Update host and edgetpu to 2.0.2, then make an int8-tflite and compile the tflite with -a option.
As it changed from 2.0.1 to 2.0.2, many ghosts disappeared.

If there are still ghosts, it might be a problem during the quantization process.

[fogelton]

I checked with the new version, reexported the model, the ghosts are still there.