I have a dataset of only 5 images that each have 60 labels on them. 30 of each class.
Roboflow created a reasonably accurate model from this dataset but YOLOv7 did not at all.
I was wondering why this might be?
As i have 150 labels of each class is this an ok dataset to start or is it more important to have more images?
Do you have a Universe link or a workspace/project ID to refer to?
Generally, it’s hard to give advice on what amount of data is enough to train a good model. It can involve some amount of trial and error, and it depends on other factors as well such as augmentations, training from checkpoints, and model type.
The project is
I guess my question is:
Is 30 labels for 5 images equivalent to 150 images with one label per image?
Not necessarily. There’s no specific rule as to
x many annotations or
x many images will do as well as something else. Machine learning models learn on patterns. If the pattern is easy for the model to recognize, it might require less training data than some other projects.
It’s best not to focus on the pure count of labels/images. Rather, try training a model with a certain amount in a dataset, then evaluate the performance and see where you need more data. Roboflow has tools to make that process easier. Before you train your model, you can use Health Check to see where your project could benefit from more data.
After you train your model. you can look in the details for your trained model too: