unit 4.5 - Labeling data
Labeling automatically
If you want to create your own dataset for localization and categorization, you can use some open-source trained neural network models to do so. Here below we show an example of what you can do.
What is the magic?
Here are some great tools (year 2024):
Text -> bounding boxes:
GroundingDINO uses BERT and SWIN to extract test and image features and has powerful zero-short accuracy, but it may be slow.
Because it was trained with a language model, it has complex capabilities for a large variety of objects, for example it can detect a “dog tail”:
It even has the capability to filter based on the relationship between objects, for example it can tell “chair” versus “men sitting on chair”:
Bounding boxes -> segmentation masks:
Segment-Anything can generate segmentation masks for all objects in the image.
It can even be used on medical images:
