Custom yolov3 detects the whole frame

Currently Deepstream 5.0 is installed (on a tx2) and the sample app for custom yolo detection runs perfectly. But when a custom trained yolov3 weights (1-class ) file is passed, the whole frame where multiple objects are present is detected as a single object. Similarly, the nvtracker runs keeping track of the whole frame.

Please provide guidance in solving this issue.
(Tried altering the net-scale-factor)


Do you delete the .engine file first?
Deepstream will load the file directly without compiling it from a new .cfg if it is presented.

For a customized model, you may need to update some parameters first.
You can find the details in the document below:

If the issue goes on, could you share the model, configures, and source(if any) with us?


Thank you for your support, it was indeed the model engine which caused the issue.