search

DanaHan / Yolov5-in-Deepstream-5.0

Describe how to use yolov5 in Deepstream 5.0
209 stars 57 forks source link

Yolov5s performances problem? #2

Closed luvwinnie closed 4 years ago

luvwinnie commented 4 years ago

It seems like the performances on example video very bad, any idea on which issues? The bounding boxes looks like not very stable.

luvwinnie commented 4 years ago

your code which seems like using the NvDsInferYoloCudaEngineGet for parsing the tensorrt engine, is this correct??

DanaHan commented 4 years ago

I have no problem. I can run yolov5 in deep stream 5.0 with fp16 model in a Xavier about 60 fps. Of course, the detect class num also influence the performance.

luvwinnie commented 4 years ago

@DanaHan I'm Glad to hear that! Maybe my model have some problem I would like to try again with your steps. Just in case, how did you run the app? Is it using the deepstream-app -c config_file?

DanaHan commented 4 years ago

LD_PRELOAD=$PWD/libmyplugin && deepstream-app -c config_file

DanaHan commented 4 years ago

your code which seems like using the NvDsInferYoloCudaEngineGet for parsing the tensorrt engine, is this correct??

I am read all source codes recently. It seems that NvDsInferYoloCudaEngineGet is called only when you use non-tensorrt engine model (ONNX,UFFT, etc). And this function seems used to build engine model. I'm not sure whether this is correct.

luvwinnie commented 4 years ago

@DanaHan I tried again with your apps it seems like even "Person" class not predicted very well, your test seems like perform very well on Person? And I have tried to replace the .cu files from this issue #3. It classify my face as "cup" and my body as diningtable

luvwinnie commented 4 years ago

@DanaHan I delete all files and followed again your step, it works well! Seems like the weight of the yolov5s.pt not really good, I redownload all the weights and tensorrtx, and your repository. it work fine !