Training custom dataset on RFDETR but retaining existing classes

Community Help
1
  • Project Type: Object Detection with RFDETR

Hi all,

I’m working with the RFDETR model for object detection and typically start training from the public checkpoint. I have a few custom object classes I’d like to detect, but I also want to retain the existing pretrained classes like person, car, etc.

Is there a recommended or best practice for combining both the pretrained classes and new custom ones?

One approach I’ve considered is running two separate models—one using the original pretrained weights for standard classes, and another trained specifically on my custom dataset. But this feels a bit clunky, and I’m wondering if there’s a more efficient or integrated solution that others have used.

Any advice or examples would be greatly appreciated!

Thanks in advance.

2

I think the current way you’re working with the models is the best/easiest way to do what you’re aiming for. A fine-tuned version of RF-DETR can only detect the classes its been trained for - meaning if you wanted one model to detect the base classes as well, you’d have to include annotations for those during training (treating it like a “new” class). But this seems like a lot of extra effort just to avoid using two models!