I've a large multi-class object detection image dataset. The goal is to use a Yolo(v11) model to be trained on the aforementioned dataset to solve the object detection task.
My intuition says that the presence of unlabeled class instances degrades the model performance of my Yolo model because those unlabeled patterns can be understood as 'training false positives' because my model could be penalized if found the unmarked object during the training phase.
In my context, I've trained a vision model that can detect the objects i want to detect with with high accuracy except that this second model just detects the target objects except for their labels.
Take as an example of my problem that my goal is to detect objects belonging to the classes $\{Cat, Dog\}$ on images with cats and dogs (and no other animals) while I have a very good model that is able to detect animals.
Question:
- I was wondering to know if it is possible to mark the objects in my dataset labeled with a generic class in order to be ignored during the Yolo training phase?
It's obvious to my that I can mask the generic labeled objects with blur or cropping them, but I'm unsure about how appropriate is this procedure or if there is a canonical procedure choice when dealing with this class of problems.
Thanks in advance!
