[Q] sample drivenet for multiple GMSL cameras on AutoChauffeur

Dear,

I configured two GMSL cameras to AutoChauffeur.

  • DriveWorks : 5.0.5.0b
  • camera 1 : SF3323(120 FoV) at group B (mask: 0001)
  • camera 2 : SF3325(60 FoV) at group A (mask: 0001)

Each camera works pretty well with sample_drivenet sample respectively and detect surroundings as expected.

% ./sample_drivenet --input-type=camera --camera-type=ar0231-rccb --csi-port=ab (or cd)

However, sample_drivenetNcameras with two cameras together does not detect objects at all.

% ./sample_drivenetNcameras --input-type=camera --camera-type=ar0231-rccb --selector-mask=00010001
  1. Video displays are OK, but no detection
  2. ar0231-rccb-bae as a camera-type tries to detect something but wrong results.

Any idea would be appreciated.
Thanks,

Dear nvdpx2,

We will fix this issue next coming DPX2 SDK. Thanks.

Hello Steve,

Then current DW_0.6.67 does not support DriveNetNCameras Sample described at DNN/Detection, doesn’t it?

How about previous version?

Thanks,

Dear nvdpx2,

That’s a DW 0.6.67 version bug.
AFAIK, this symptom doesn’t happen on previous version. Thanks.

Thanks Steve,

Hi Steve,

I am also trying to get the sample_drivenetNcameras application working with multiple cameras; I too can confirm that no objects are properly being detected and tracked in DW 0.6.67.

When you mentioned it is a bug, did you mean that it is the bug with how the sample_drivenetNcameras is using the DriveWorks API, or did you mean it is a bug in the library? If it is the former, I suppose I can modify the sample_drivenet application to support multiple cameras.

Also, from the “DriveWorks SDK Reference” document, it is mentioned that “The sample does not use foveal detection, hence the quality might be worse then in a normal drivenet sample. However, the sample demonstrates how to feed multiple images at once to the DriveNet. Since DriveNet supports a batch of 2, we are doubling the throughput of the images processed by the network, by taking a hit in detection quality. In order to achieve best detection quality, refer to foveal detection in sample_drivenet.”. Is this the answer on why the sample_drivenetNcameras application doesn’t detect and track objects properly, rather than a bug?

Thank you for looking into this.