Train_ssd.py Annotations (must they be rectangular?)

[ll16j33m]

Hi,
I’ve been training mobilenet on datasets produced by VoTT. The training has been successful thanks to thorough documentation and great support by dusty-nv and others. My question is, am I wasting my time by annotating images with polygons of 5+ vertices? The images I am labelling for object detection are typically circular in shape, but I’m not sure if perhaps the script is transforming my regions into rectangles anyway.

Any information is greatly appreciated, thanks.
Best,
Jacques

[dusty_nv]

Hi @ll16j33m, train.py and SSD-Mobilenet only work on bounding boxes (rectangles), so the polygon annotations are being converted to boxes.

That’s not to say that annotating your data with polygons isn’t useful for the future - if you decide to implement or use a DNN that handles polygons natively (although I personally haven’t seen much of these other than 6-DoF 3D bounding boxes). However if you have no intentions of this than yes, you can save yourself time by just annotating with bounding boxes.

[ll16j33m]

Hi Dusty, thank you very much for your quick and informative reply!