dbolya
commented
4 years ago That comment was not meant as "if we don't generate boxes, we'll do better", it was meant as "our bbox detector is not state of the art, meaning it pulls the rest of the method's performance down with it". If you wanted to try to improve that, I'd suggest trying to implement a better object detector in the first place, not removing bboxes entirely (which would make the method worse, actually).
Hi, thanks for your landmark work that can inspire me. In the Understanding the AP Gap, you wrote that "which suggests that the gap between the two methods lies in the relatively poor performance of our detector and not in our approach to generating masks". Therefore, I want to train the model without using annotations of bounding box on the Pascal 2012. In other words, I hope model only can generate masks , not both mask and bounding box. Can you give mo some advices or tips? Thanks again.