As per my understanding, ML works on the concept of the type of Input+Output you feed during training, will expect(Input) and predict the same output.
So labeling is must for Supervised learning.
For Image localization use case, you have to trained your model with similar type of data i.e. label should have full annotation/details with respect to dimensions(x,y,height,width) and object class. Only then you can expect the Object localization(bounding box) during prediction.
Few of my posts possibly be helpful for you:
How YOLO training and prediction works for an object fall in multiple grid?
YOLO: How many bounding boxes?
Please go through these videos for more details. Object Detetction and Localization