Detectnet : DIGITS inference output inconsistency with tx1 inference result

HI,

I modified the detectnet protocol by reducing grid space from 16 to 8, In order to detect small objects in a frame. the DIGITS inference results is OK by detecting 21 small bboxes(e.g. 9x7, 7x4, 10x5, 7x10 etc.), so I deployed the caffeModel to my tx1 to execute on-line inference. However the GIE engine could not re-produce the same bboxes as was seen on DIGITs. Instead, it pedicted many big bboxes(e.g. 100x125)

Could you tell me what could be the reason?

regards
Hoover

Hi,

Could you share how do you inference DetectNet on Jetson first?
Do you use the NvCaffe or Jetson inference?

Thanks.

Hi, AastaLLL,

I used Jetson inference.

Hi,

Any updates?

Hi,

1) Please check if any update is required for ow/oh, cell_width/cell_height and scale_x/scale_y
https://github.com/dusty-nv/jetson-inference/blob/master/detectNet.cpp#L522

2) Please check if any update is needed for the mergeRect function:
https://github.com/dusty-nv/jetson-inference/blob/master/detectNet.cpp#L450

Thanks.

  1. Please check if any update is required for ow/oh, cell_width/cell_height and scale_x/scale_y
    https://github.com/dusty-nv/jetson-inference/blob/master/detectNet.cpp#L522

—no updates seems to be necessary, as I input 640x540 image, and saw the ow=80, oh=67, cell_width=8, cell_height=8, scale_x =1, scale_y=1;

  1. Please check if any update is needed for the mergeRect function:
    https://github.com/dusty-nv/jetson-inference/blob/master/detectNet.cpp#L450

—It is not likely related to this clustering function. Before “mergeRect”, I saved all bbox with coverage>threashold in a file in KITTI label formats, then I viewed them with ALP’s labeling tool. They are already as big as,say 100x100.
i.e. I saved and reviewed bboxes before clustering at https://github.com/dusty-nv/jetson-inference/blob/master/detectNet.cpp#L567

Thanks
Image_0.txt (83.7 KB)

Hi,

In the attached Image_0.txt are the 997 bboxes before mergeRect(), they don’t look like can be clustered into small bboxes at all. What might be your suggestion?

Thanks

Hi, AastaLLL,

Please let me know if your team need me to provide further information in order to analyze this problem.

Best Regards,

Hoover

Hi,

Another common issue is about the mean subtraction.
Do you have any scale or subtract layer inside your model? Ex. power layer

Please make sure the data range in jetson_inference is same as the one you used for DIGITs first.
https://github.com/dusty-nv/jetson-inference/blob/master/detectNet.cpp#L491

Thanks.

Not sure if this helps, but I’ve successfully installed NVIDIA DIGITS on Ubuntu 18.04 LTS operating system, then was able to following the tutorial for DetectNet for the Jetson Xavier.

Here is the DetectNet Tutorial (scroll down):
https://github.com/dusty-nv/jetson-inference#system-setup

Here is a video walkthrough of getting NVIDIA DIGITS natively installed with Ubuntu 18.04 LTS:
https://cudaeducation.com/nvidiadigits/

Hope this helps.

-Cuda Education
cudaeducation.com