I am wondering if it is required or not that the masks used for binary instance segmentation are complete.

For instance, I want to find the buildings in aerial imagery. If my masks cover, let say between 25% and 75% of the training image buildings, can these samples be used for training the model? Does the masks should cover 100% of the buildings in the samples?

I know it would be better to mask all buildings, but if the mask is not complete, can it be used for training or not?

I am using a deeplabv3 model.

Here an example of what I mean enter image description here


1 Answer 1


You can train a model with that dataset, but it will perform worse compared to a higher fidelity dataset. Training on images with unannotated buildings is essentially instructing the model to stochastically ignore buildings, which is confusing.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .