COCO checkpoint can't detect COCO classes

Community Help
1

I have a couple questions about transfer learning.

  1. I trained a new Yolo11 fast model and chose COCO as the starting checkpoint. When I test the model with say a person, they are not detected, even though that is a class within COCO. Am I miss understanding something? I expect to be able to detect classes within COCO dataset if I use a YOLO model trained on COCO as a starting point. Note: I did this all from the Roboflow web interface.

  2. How can I extend the learning of a class in the COCO dataset. For example. The COCO dataset is trained on birds. Can I train a new model on Roboflow, use COCO model as the starting point and extend training for bird class with new images of my own?

Thanks.

2

Hi @duck-rep!
Great questions! To address them in order:

  1. If you want your model to detect a person, you must also have a “person” class within your dataset. When you use COCO as the starting checkpoint, you’re using those pre-trained weights for transfer learning. The final model architecture is rebuilt to match your dataset’s classes.
  2. Yes, if you have a “bird” class in your dataset, your model will benefit from the COCO pre-training on birds (since COCO includes “bird”). This means your “bird” class will likely train faster and perform better than training from scratch.
1 Like
3

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.