The Pytorch Faster-RCNN model returns bounding boxes, labels and scores.
I’m not sure, are the scores represent how likely is the bounding box to contain an object (objectness score) or classification score?
1 Like
In the forward pass of the RPN module (which is part of FCRNN module) there is an objectness variable that holds the objectness score but it’s not stored as part of the class. I’m not sure if those are the same scores you get when inference the model.
Maybe @ptrblck can help 
.
1 Like
By looking at the code of the PyTorch implementation of Faster RCNN and especially roi_heads.py, we can find the line
pred_scores = F.softmax(class_logits, -1)
in the function postprocess_detections which returns boxes, labels and scores,
which means scores are the classification scores.
1 Like