One thing I’m still unsure about is whether we are returning Bounding Boxes or Masks over the objects in order to calculate AP and AR. The datasets themselves seem to have both.
Thank you,
William
One thing I’m still unsure about is whether we are returning Bounding Boxes or Masks over the objects in order to calculate AP and AR. The datasets themselves seem to have both.
Thank you,
William
@william_locke : This challenge focussed on the instance segmentation task, and hence your submitted code has to return the polygons representing each of the detected instances. Hope, that clarifies your question.
The official baseline code for MMDetection and Dectectron2 uses both bbox
and segm
for evaluation on the validation dataset. You can change the behavior very easily.