How to avoid multiple (overlapping) bounding boxes

[rajeshroy402]

Hey @dusty-nv , thanks for this awesome repo and sharing your content with the community. I retrained a model on xavier nx using the ssd object detection model by following your YT tutorial. I wanted to share a concern that I am getting overlapped bounding boxes with both higher then 80% confidence and this is something I don't need. Can you tell me which python file will I have to edit and what exactly changes do I need to perform?

Yoiu shared an easy method to retrain using mobilenet ssd but what if I want to use Yolov3. do I need to start from scratch? Alongside, can I use the onnx model with DS-5.1? Is there any sample repo which replicates the same?

Thanks,

[dusty-nv]

Hi @rajeshroy402 see this comment to decrease the clustering threshold:

https://github.com/dusty-nv/jetson-inference/issues/880#issuecomment-754885337

Yoiu shared an easy method to retrain using mobilenet ssd but what if I want to use Yolov3. do I need to start from scratch?

I don't support YOLOv3 in this project, but here is a tutorial that deploys YOLOv3 using TensorRT: https://jkjung-avt.github.io/tensorrt-yolov3/

Alongside, can I use the onnx model with DS-5.1? Is there any sample repo which replicates the same?

Here is the post-processing bounding box parser for using SSD-Mobilenet ONNX with DeepStream: https://elinux.org/index.php?title=Jetson/L4T/TRT_Customized_Example#Custom_Parser_for_SSD-MobileNet_Trained_by_Jetson-inference

[rajeshroy402]

Thanks for sharing these resources @dusty-nv . Wanted to know like which file should be edited if I want to apply some additional logic on my detectnet custom trained model?

And is it possible to run the custom trained model on two or more rtsp streams at one time? If yes, then how. Thanks!