Questions for "Detecting Twenty-thousand Classes using Image-level Supervision"

-----------------------------------------------------------------------------------------------

Please send your answers to: bravoma@cs.uni-freiburg.de by 13:15 on 03.06.2022

1. Describe how does Detic loss improves object detection on novel categories. (~3 sentences)

2. Detic uses the max-size box (Table 7) for Detic loss. Why do you think this variant performed the best? (~2 sentences)

3. How or what would you add to Detic model to improve it's performance? Write at least 2 ideas and tell why you think they would help. (~4 sentences)