charlesCXK / RGBD_Semantic_Segmentation_PyTorch

[ECCV 2020] PyTorch Implementation of some RGBD Semantic Segmentation models.
MIT License
301 stars 43 forks source link

Little question #5

Closed TXH-mercury closed 4 years ago

TXH-mercury commented 4 years ago

Hi, Thank you for your great work for us ! Have you used multigrid as in the code to get the 52.4mIOU on the NYUD dataset?

charlesCXK commented 4 years ago

Yes.

TXH-mercury commented 4 years ago

@charlesCXK Thanks! Noticing the relatively big performance difference between NYUDv2 Datasets and SUNRGBD datasets ,I have also encounter this kind of big performace gap on those two datasets , But some methods like ACnet shows no explict performance gap . Have you figured out some reasons?

charlesCXK commented 4 years ago

Hi, I think there are two possible reasons.

  1. Maybe we didn't use the suitable training setting for SUNRGBD, such as lr schedule, etc.
  2. SUNRGBD has many wrong labels (as shown in our paper), while the labels of NYUDv2 are clean enough.

Unfortunately, we haven't figure out how to improve the performance on SUNRGBD (including baseline), but you could check their code and maybe you could find some tricks we ignored before.

TXH-mercury commented 4 years ago

All rights , very helpful , I prepare to try tuning some training hyperparameters first. Thanks very much .