Dear Organizer,
It might be a fault on my side, but I noticed that in **eval.py** in official github, the same threshold is used across all frames, while in the DeepLab baseline evaluation (**utils/trainer.py**), the threshold seems to be determined per batch, resulting in a higher AP score than it should be. Since this code will be used to train models for Task 2?s evaluation, will this issue be corrected in the official evaluation?
Additionally, could you publish the baseline score using the pre-trained weights from DeepLab and DexiNed? The predictions made with the provided weight are less accurate than expected, making it difficult to confirm whether the official model has been accurately reproduced.
Sincerely,
Created by Shunsuke Kikuchi ne_gi_chi__ Awesome, thanks for the immediate response! scores on validation have been posted Thanks. Will the baseline scores be released by the deadline?
best,
Hi @ne_gi_chi__
The script eval.py in the official Github is the correct code. The evaluation code found in our DeepLab baseline is outdated and we updated it now on Github.
As for publishing the baseline scores, we will publish those soon as well.
Best regards,
Rema Daher