This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA4VM: TIDL inference of YOLOV3 use AR0820 camera

Part Number: TDA4VM


Hi TI members:

I used AR0820 camera with 3840x2160 resolution as source and run YOLOV3 to do inference. Bad result I got.

I save the output of scaler node to NV12 image and I found this image is wrong.

Is scaler node unable to resize image from 3840x2160 to 416x416?

Are there any suggestion to fix this?

Thanks.

YouJun Shih

  • Hi YouJun Shih,

    Which scalar node are you using? 

    Scalar, as such, supports downscaling upto 1/4x only. But i think few scalar nodes supports downscaling by any resolution, by running it multiple times. 

    Regards,

    Brijesh

  • Hi, Brijesh,

    I used app_tidl_od_cam to do inference, the scalar node should be VPAC Multi-scaler.

    How to run scalar node multiple times? Can you give me some suggestions to achieve this ?

    Thanks.

    YouJun Shih

  • Hi YouJun Shih,

    I think multi-scale does not support it. So can you first downscale the image by 4 using multi-scaler and then use scalar node to do further downscaling?

    Regards,

    Brijesh

  • Hi, Brijesh,

    I think that app_tidl_od_cam use multi-scaler as scalar node. So I can't understand the meaning of "first downscale the image by 4 using multi-scaler and then use scalar node to do further downscaling". What's difference between muti-scaler and scalar node? Do you mean to create two scaler node to do downscaling multiple times?

    Thanks.

    YouJun Shih

  • Hi YouJun Shih,

    There are different types of scalar nodes supported using the same HW. 

    Multi-Scalar Node: This node takes one input and generates at max 5 outputs from the same input. 

    Scalar Node: This node takes one input and generates one output. 

    I am not sure how app_tidl_od_cam useecase uses scalar. But if it uses multi-scale node to generate 5 outputs, then you use multi-scale node to generate output up[t 1/4x resolution and then use scalar node on the output for further scaling. 

    Regards,

    Brijesh

  • Hi Brijesh,

    Thanks for your support. I will apply your suggestions to app_tidl_od_cam to check the result.

    Thanks.

    YouJun Shih