Deepstrea5 /ONNX problems Bounding box shifted

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hello,

Please, I have a problem that is blocking me and I can’t move forward,
I created a personalized model with this tutorial PyLessons then convert the weights to ONNX format for use with deepstream5, I used this directory to interface deepstream with onnx https : //github.com/thatbrguy/Deep-Stream-ONNX, but my prob lem is that the bounding boxes are shifted and I can’t find the right parameters to solve this problem,
Any suggestions please ??

image_save_3
image_save_3960×540 151 KB

Thanks

3

Hi,

Before converting to Deepstream, could you get the expected result with ONNX format?
If yes, could you share the onnx model and the source for reproducing onnx and Deepstream output with us?

Thanks.

4

I did not do the test with onnx without deepstream, on the other hand I took the weights generated from the training (before converting them to onnx), I did a little test in python I had the expected results ,

source : const char * streamModel = "uridecodebin uri=file:///home/sylia/DeepStream-Gstreamer/stream/video1.mp4 "
"! nvvideoconvert "
"! m.sink_0 nvstreammux batch-size=1 width=960 height=540 name=m "

    "! nvinfer  config-file-path=configs/inference.txt "
    "! nvtracker ll-lib-file=/opt/nvidia/deepstream/deepstream-5.0/lib/libnvds_mot_klt.so  tracker-width=960 tracker-height=540 "
    "! nvdsanalytics name=nvanalytic0 config-file=configs/analytics.txt "
    "! nvdsosd name=osd " 
   // "! tee name=t "
    //"t. "
    //"! queue name=screen " 
    "! nvvideoconvert  "
    "! ximagesink " ;

inference file : inference.txt (2.2 KB)
analytics file :analytics.txt (2.8 KB)
yolo library yolo_custom_bbox_parser.tar.xz (32.0 KB)

5

onnx model model.onnx - Google Drive

6

Hi,

Thanks for your sharing.

We are going to reproduce this issue internally.
Will share more information with you later.

7

thanks

8

Hi,

We try to reproduce this issue but failed to download the ONNX model due to permission.
Could you help to enable it?

Thanks.

11

Try this please

12

Thanks.
We can download the file correctly.
Will share more information with you later later.

13

Thanks

14

Hi,

Thanks for your patience.

We can reproduce this issue in our environment.
It seems that the model has it’s own output parser rather than the standard YOLO one.

May I know more detail about the training source.
Could you get the correct result with the standard parser from GitHub - pjreddie/darknet: Convolutional Neural Networks?

Thanks.

1 Like
15

Hi,

Here is a similar issue for your reference:
https://github.com/NVIDIA-AI-IOT/deepstream_triton_model_deploy/issues/8

Thanks.

1 Like
16

Thanks