NMS for different classes

Hi. I trained the model with my own dataset and it's working fine. During inference, some objects are filtered out because of low confidence. When I lower the DETECTION_MIN_CONFIDENCE parameter, the remaining objects are indeed detected, but sometimes overlapping bounding boxes/masks are located in the same object in the image. These overlapping masks are not filtered by NMS because they're from different classes (labels), i.e., the network is predicting the same object, with almost exactly the same mask, but assigning different labels to those predictions.

I was wondering if there is a way to apply NMS for different classes too (in my case, it is rare to find overlapping objects of different classes in the ground truth images).

Start by locating the inference path that uses DETECTION_MIN_CONFIDENCE and applies NMS. Trace how class labels affect filtering, then determine how cross-class overlap should be handled without removing valid distinct objects. Done means overlapping predictions for different labels can be filtered according to the requested behavior, with coverage for this case.

Written by the indexing model from the issue text.