I am doing a proof-of-concept thing to see if Mask RCNN can do a good instance segmentation on my own dataset.

The issue is that I have to annotate the data myself and it takes long time to annotate one single sample, so I would like to test if this idea will work with a lower cost. My workaround so far is to annotate a minibatch (~16 images as training samples, no validation or test set is used) and feed the minibatch on the model to see if the model can overfit to the minibatch.

For sake of cost, I would like to infer the model performance on generalized unseen data with such a minibatch. Suppose that the model can really overfit to the minibatch, can I come to the following conclusion?:

"the model has enough capacity to capture the complexity of the data pattern, with enough annotated training samples, the model can finally generalize well in performance"
^^true or false?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.