search

Daniil-Osokin / lightweight-human-pose-estimation-3d-demo.pytorch

Real-time 3D multi-person pose estimation demo in PyTorch. OpenVINO backend can be used for fast inference on CPU.
Apache License 2.0
653 stars 137 forks source link

Question about TensorTR on Jetson nano #78

Closed chris666-sys closed 2 years ago

chris666-sys commented 2 years ago

@Daniil-Osokin Hello, first of all thanks for your github Work

I ran this project on jetson Nano using TensorTR

Currently I can use python scripts/convert_to_trt.py --checkpoint-path human-pose-estimation-3d.pth --height 256 --width 256 command to get human-pose-estimation-3d-trt.pth

According to the document, I set the network size --height 256 --width 256, but when I ran the command python demo.py --model human-pose-estimation-3d-trt.pth --use-tensorrt --video 0, the human posture was not detected

So I try to change the network size --height 192 --width 192, (--height 32 --width 32) and pass --height-size 192 to the demo.py ，still can't detect the human posture

I referred to the discussion #41, but human posture was never detected with TensorTR ( human posture was detected without TensorTR)

How should this problem be solved? Thank you for your reply...

Daniil-Osokin commented 2 years ago

Hi! Which size has network input (you can get it with print(scaled_img.shape) in the demo.py)?

chris666-sys commented 2 years ago

@Daniil-Osokin Hello, according to your guidance, I solved this problem, the current network size is (--height 256 -- width 336), can now show the human posture, but it is easier to identify other items into the human body, how should I adjust the network parameters to solve this problem? Thank you very much...

Daniil-Osokin commented 2 years ago

To have better quality, use bigger input resolution (e.g. multiply current height and width by 1.5).

chris666-sys commented 2 years ago

@Daniil-Osokin Hello, thank you for your reply.

I modified the network size and multiplied the height and width by 1.5 times, Then run the demo.py, but got the following error

QQ图片20210727114338

Then I tried to multiply the height and width by 2 times and got the same error

It seems that the size of this network cannot be changed...Thank you!

Daniil-Osokin commented 2 years ago

If you have reconverted the trt model with new height and width values and see this error, then it is possibly due to lack of free memory. So this trick will not help, but you can try it on the desktop GPU to see if quality will increase for your use case.

chris666-sys commented 2 years ago

@Daniil-Osokin Okay, thank you for your reply, thank you very much！