use multi datsets to train YOLO-World

[pixiaopi2018]

Search before asking

• [X] I have searched the Ultralytics YOLO issues and discussions and found no similar questions.

Question

Hello, I want to use my own datasets to train a YOLO-World detector. But there is a question, for example, dataset_A has car/person in images, but only 'car' is labeled; datset_B also has car/person in images, but only 'person' is labeled. Thus, how can I put dataset_A and B to train YOLO-World simultaneously, the unlabeled objects won't affect the learning ?

Additional

No response

[Y-T-G]

You need to have both labeled in all the datasets.

[pixiaopi2018]

You need to have both labeled in all the datasets.

Thanks for reply! I agree with you when training normal detector, but this is a ovd detector, the distribution of its train data is very wide (like Flickr/GQA/Object365). For example, some categories in Flickr, must not be labeled in obj365. So I have this question.

[glenn-jocher]

@pixiaopi2018 thank you for the clarification. For YOLO-World, you can indeed train with datasets where not all objects are labeled. The model is designed to handle such scenarios effectively. Just ensure that your datasets are properly formatted and the labels are consistent across them.

[pixiaopi2018]

@pixiaopi2018 thank you for the clarification. For YOLO-World, you can indeed train with datasets where not all objects are labeled. The model is designed to handle such scenarios effectively. Just ensure that your datasets are properly formatted and the labels are consistent across them.

Thanks! In yoloword, what design is made to do this (same category labeled/not labeled in diffrerent datasets, but datasets can be trained together)? And what "the labels are consistent across them" means? Thanks for your patience agian!

[glenn-jocher]

YOLO-World uses a "prompt-then-detect" strategy, allowing it to handle datasets with varying labeled categories. By "consistent labels," I mean ensuring that the same object categories have the same label names across all datasets. This helps the model understand and detect objects correctly during training.