Deepstrea5 /ONNX problems

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hello,

please, after succeeded to having my custom model in onnx I did the tracing with deepstream5, but the detection results are not satisfactory, this is the deepstream / onnx test output

image_save8
image_save8960×540 132 KB

besides I did a test in python with the weights generated during training (yolov3-custum.index / yolov3-custum.data-00000-of-00001)
the results are better as the following picture shows

image2_sav
image2_sav1300×953 388 KB

Does the conversion reduced performance ?
Thanks

3

Hi,

This issue may occur due to the incorrect parameter.
Suppose you are using a customized YOLO model, please check below document to update the parameter:
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_custom_YOLO.html

Thanks.

4

thank you for your answer, I always have offset bounding boxes …

I use onnx model, I managed to make the onnx model / deepstream5 interface, on the other hand I cannot find the right parameters to avoid the declining of the bounding boxes

Onnx / deepstream interface I used this directory: GitHub - thatbrguy/Deep-Stream-ONNX: How to deploy ONNX models using DeepStream on Jetson Nano

5

Hi,

Sorry for the late update.
Let’s track the following progress on Deepstrea5 /ONNX problems Bounding box shifted directly.

Thanks.

6

Thanks