Advanced labeling techniques for object detection

Hello! Are there any ways of using advanced labelling techniques to create a dataset for image object detection i.e. create bounding boxes + object label automatically? If that is the case, which ones are the most commonly used?

Thank you

Hi @Matias_Saez

well, if you are talking about a supervised learning approach, you need to provide the ground truth, in other words the label (and BBR) for the training set. If you think about it, it is the knowledge that you’re injecting inside the model.
In general, you could try with a semi-supervised approach. generating label automatically. But you need a good way to do that. If you have already a method to generate perfect BBR + label then you don’t need to train a model, you already have it.
One could think to use a model already trained on ImageNet (one that you found in every library TF or PyTorch.
But image classification, the kind of problem addressed in ImageNet, assume only one object in the image, or label only the main object. In Object detection you have many object, therefore you should in advance segment in many smaller rectangles in order to have a single object inside. Again, if you have something for doing that, why you need to train a model.