Azure-Kinect-Sensor-SDK: Body tracking k4abt_simple_3d_viewer crashes on start up
Hello everybody, I am trying to set up the body tracking abilities that the Azure Kinect is supposed to have. Hereafter my PC configuration and the drivers installed:
OS: Ubuntu 20.04 CPU: Intel i7-1050H CPU 2,60 GHz GPU: Gtx 1050 GPU driver and Cuda version (output of nvidia-smi): NVIDIA-SMI 495.29.05 Driver Version: 495.29.05 CUDA Version: 11.5
Camera settings
- SDK Version: 1.4
- Body tracking Version: 1.1
Note at this time the Cuda version was cuda10.x I have installed the latest body tracking packages:
- libk4abt1.1_1.1.0_amd64.deb
- libk4abt1.1-dev_1.1.0_amd64.deb
with the following commands:
sudo dpkg -i libk4abt1.1-dev_1.1.0_amd64.deb
sudo dpkg -i libk4abt1.1_1.1.0_amd64.deb
After that, I was able to run the k4abt_simple_3d_viewer from the command line (even if the computations were running on CPU) wit the following warning:
2021-11-06 18:06:27.440808922 [W:onnxruntime:K4ABT_ORT_Environment, cuda_execution_provider.cc:1885 GetCapability] CUDA kernel not found in registries for Op type: Sign node name: Sign 2021-11-06 18:06:27.442896633 [W:onnxruntime:K4ABT_ORT_Environment, fallback_cpu_capability.h:140 GetCpuPreferedNodes] Force fallback to CPU execution for node: Shape__40
After that, I supposed that the problem was the Cuda version so I decided to upgrade it to the latest one (the one that I actually have): Cuda 11.5. (I am aware from other threads that the tested Cuda version with the body tracking is the Cuda11.1.1 but because of some Kernel problems I am not able to use that version and I don’t want to downgrade my Kernel)
k4aviewer gives me absolutely no problems in running, all the cameras are perfectly working.
Here is the output as soon as I run from the terminal k4abt_simple_3d_viewer
2021-11-06 18:12:43.229226675 [W:onnxruntime:K4ABT_ORT_Environment, cuda_execution_provider.cc:1885 GetCapability] CUDA kernel not found in registries for Op 2021-11-06 18:12:43.231439200 [W:onnxruntime:K4ABT_ORT_Environment, fallback_cpu_capability.h:140 GetCpuPreferedNodes] Force fallback to CPU execution for no *** buffer overflow detected ***: terminated Aborted (core dumped)
No other logs are available. So as you can see still the camera is not able to run with Cuda and moreover now does not even start anymore. Any help will be very appreciated.
Niccolò
About this issue
- Original URL
- State: closed
- Created 3 years ago
- Comments: 18
downgrading is not a sustainable solution as you can’t downgrade the newer graphics cards to 470 (i have a 3080ti, and this isn’t an option for me)
This issue can be resolved by #1696 and #1555. It seems like microsoft didn’t make the adaption bewteen new vision driver ( Higher than 472 ) and SDK. I found the nvidia driver you used is 495. What you need to just degrade your driver. I have the same probem few days ago, but I degrade driver to 470 version and it solve my problem. And I wish microsoft make the adaption bewteen new vision driver and SDK quickly.
That did fix the issue and it works now. Thanks!
Seriously need a proper fix for this, that isn’t downgrading to a 6 month old driver. It has been reported in multiple tickets for 6 months, and I need to keep my nvidia drivers up to date for lots of other software.