bytedance
About the question of Schedule
Hello, author. Thank you very much for your open-source contribution. I noticed that the schedule settings in the paper are all 6x epochs. Why was such a long epoch chosen? Faster R-CNN would likely see minimal performance improvement with such a setting. Additionally, I noticed that the DINO configuration you open-sourced is actually 1x. I am not sure why. Sorry for the interruption.
Hi, thank you for the attention.
-
About the schedule, as mentioned in the paper(Fig. 5), we notice that traditional schedule setting up to 3× schedule gradually improves the performance, while further scaling leads to overfitting. However, scaling data with DE as well as the schedule further improves the performance, indicating that DE is an effective data scaling-up technology rather than a simple data replay.
-
About the configuration, sorry for the confusion, the code is updated and please refer to here for more information.
