search

clovaai / ECLIPSE

(CVPR 2024) ECLIPSE: Efficient Continual Learning in Panoptic Segmentation with Visual Prompt Tuning
Other
41 stars 5 forks source link

Possible problems and some questions about the paper #6

Closed ZhuYuChenNO1 closed 4 months ago

ZhuYuChenNO1 commented 4 months ago

Hello! Thank you for your great work. I have some questions about your paper.

  1. In Table 1, For this row of FT data, it seems your data comes from CoMFormer. However, can 100-5 (11 tasks) and 100-50 (2 tasks) be written in the wrong positions?
  2. In Table 1, Regarding the performance of CoMFormer, why is the data for 100-50 (2 tasks) inconsistent with the original paper?
qjadud1994 commented 4 months ago
  1. Oh! Thank you for pointing out. It is an error as you mentioned.
  2. We successfully reproduced the performance of CoMFormer for all scenarios, except only for 100-50 (2 tasks). We tried to reproduce the performance of CoMFormer, PLOP, and MiB on the 100-50 scenario by changing various hyperparameters, but we failed. Therefore, we reported our reproduced performances of these methods on the 100-50 scenario.
ZhuYuChenNO1 commented 4 months ago
  1. Oh! Thank you for pointing out. It is an error as you mentioned.
  2. We successfully reproduced the performance of CoMFormer for all scenarios, except only for 100-50 (2 tasks). We tried to reproduce the performance of CoMFormer, PLOP, and MiB on the 100-50 scenario by changing various hyperparameters, but we failed. Therefore, we reported our reproduced performances of these methods on the 100-50 scenario.

Thank you for getting back to me. My question has been solved. As a side note,Hope you can find the error cause the fail of reproduce your ADE20K-Semantic 100-50. However, this matter is not within the scope of discussion for this issue. So, I will close this issue.