Can we use detectNet for Polygon label?

Hi @Morganh ,

I want to train one detector model which can detect barcode and want to get the corner point of that (Polygon). Please refer image attached for more clarity.

Can detectNet-V2 will resolve the problem ? or I should use MaskRcnn ?
If maskRcnn then which annotation tool can give us the Final Json Annotation file that is used in the Tao/TLT MaskRcnn sample .


We can consider the barcode as one kind of projects.
You can use an object detections network. You can try YOLOv4 or detectnet_v2 network.

Thanks @Morganh for the reply. But How can I pass the corner point there ?
Because Kitti format expect xmin,ymin,xmax,ymax but I have 8 values (X1,Y1,X2,Y2,X3,Y3,X4,Y4) which are forming polygon whose X(i),X(i+1) line is not parallel to the X axis and Y(i),Y(i+1) line is not parallel to the Y axis.
Sample image of polygon type area is mentioned above in the images.

Can you please help me out how can I get prepare the label for this problem if I have to use DetectNet-V2 ?

OK. I will check further. I misunderstand your case.

There is FPEnet in current release. And it will be extended to a generic version to detect keypoints. This fpenet-generic model can support for your case.
But it is planned for next release. Please contact Nvidia PM if you want to use it now.

Thanks @Morganh
Can you please guide me how to connect with Nvidia if I want to use generic version of keypoint Detection/Regression ? How can I get it before release and can train it using TLT/TAO for my use case ?


Please find info in Contact Us: Americas Locations & Regional Offices | NVIDIA .

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.