Opencv Dnn yolov4-tiny slower yolov3-tiny (ARM)

[sealedtx]

Hi @AlexeyAB I am trying to run tiny models using opencv dnn on my armv7 system. I successfully used opencv 4.3+yolov3-tiny-3l (before yolov4-tiny release) and it's inference time was ~1.5s. But opencv 4.4+yolov4-tiny gave me ~4.5s inference time which is surprisingly a lot. I have also tested these models on PC but yolov4-tiny is faster as excepted. What is the problem with arm? Is there some operatation included in yolov4-tiny that are much slower on ARM systems? Thank you! Notes:

• Opencv 4.3 and 4.4 are compiled with same options and flags.
• I measure time only for net.forward(), without pre-post-processing
• Both models are 416x416

[AlexeyAB]

Can you compare yolov4-tiny vs yolov3-tiny with the same resolution 416x416, the same OpenCV 4.4.0 + ARM with the same code, 1000 iterations of net.forward() in a loop?

Also try to use this code: https://gist.github.com/YashasSamaga/48bdb167303e10f4d07b754888ddbdcf

[sealedtx]

Thank you, this benchmark helped me undestand that yolov3-tiny model was pruned and this is why it overperformed yolov4-tiny.

[ghost]

@sealedtx so were you able to improve tiney v4 speed on arm?

[sealedtx]

@deepseek I didn't focus on this, performance was ~same as yolov3-tiny model, which was enough for me.