Faster R-CNN output adaption


I was using the Faster R-CNN model from the torchvision library for a transfer learning task. Everything works perfectly and I get good results. But I’m wondering about the outputs of the whole model.

The output for inference is currently

  • boxes [N, 4] with the bounding box coordinates
  • labels [N] with the predicted labels for each image
  • scores [N] with the scores or each prediction

There is currently no interface investigate the class probabilities for the classes, that are not the most probable class prediction. This information would be pretty interesting for me to get a deeper insight for false positive predictions or wrong classifications.

Is there already a preferred way of doing this? If not, I would probably suggest this as a feature in Github.

Kind regards