Sample_dnn_tensor on Drive Orin, the results are incorrect

Please provide the following info (tick the boxes after creating this topic):
Software Version
DRIVE OS 6.0.6
DRIVE OS 6.0.5
DRIVE OS 6.0.4 (rev. 1)

Target Operating System

Hardware Platform
DRIVE AGX Orin Developer Kit (940-63710-0010-300)
DRIVE AGX Orin Developer Kit (940-63710-0010-200)
DRIVE AGX Orin Developer Kit (940-63710-0010-100)
DRIVE AGX Orin Developer Kit (940-63710-0010-D00)
DRIVE AGX Orin Developer Kit (940-63710-0010-C00)
DRIVE AGX Orin Developer Kit (not sure its number)

SDK Manager Version

Host Machine Version
native Ubuntu Linux 20.04 Host installed with SDK Manager
native Ubuntu Linux 20.04 Host installed with DRIVE OS Docker Containers
native Ubuntu Linux 18.04 Host installed with DRIVE OS Docker Containers

When executing the sample_dnn_tensor located in /usr/local/driveworks/samples/bin, the correct results are not rendered.

Furthermore, the message “DNN: run infer internal batchsize 1 is not needed after” is being printed to the console.

models provided as samples are not suitable for Drive Orin?

Could you share steps to reproduce the issue? Did you verify on host or target?

Dear @SivaRamaKrishnaNV

The application was executed on the target (Drive Orin).

There are no special step required.

  1. Power on Drive Orin.
  2. “$cd /usr/local/driveworks/samples/bin” in the console of Drive Orin.
  3. Execute “$ ./sample_dnn_tensor”

I also got the exact same results, what I did was compiling a yolo v3 model to a tensor RT bin file, but I still got no detection, I have been working with this problem almost for 2 months.

Dear @HIENG_NS ,
We notice this issue in DW 5.10 release(Mentioned as bug #3931886 in release notes) and is fixed in later DW release.

Hi @SivaRamaKrishnaNV ,
We encountered this issue as well when running on host (DW 5.10, DRIVE 6.0.6, Ubuntu 20.04) when running dnn samples: sample_dnn_tensor, sample_object_detector_tracker.
Same issue arise when building from source and also running the bin files you provided.

Done some debugging:

  1. After editing the ROI (m_detectionRegion.width, m_detectionRegion.height, m_detectionRegion.x, m_detectionRegion.y) to be valid (correct yellow box is displayed on screen) there are still no detections displayed.
  2. Noticed the dwBlobSize struct m_networkInputDimensions values were quite bizarre: height=1284, width=3, channels=384, batchsize=1. Could it be that width and channels got mixed? There’s no info regarding which model is used…
  3. I examined the confidence of the model output in interpretOutput function of the example and it seems that the confidence is close to zero for all grid cells, resulting in no detection crossing the predefined COVERAGE_THRESHOLD=0.6, thus no bboxes are displayed on the image plain.

My questions:

  1. Could it be there’s an issue with the model itself?
  2. Can you share when will the fix be available?
  3. Any hints you can provide to run this sample / other sample you can share for instance over a YOLO model.


I’m also curious about the answer to alon.eshel’s question.
Specifically, I would like to know if the problem lies with the model itself or with the sample source.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.