search

jd730 / EOPSN

[CVPR 2021] Exemplar-Based Open-Set Panoptic Segmentation Network (EOPSN)
Other
52 stars 7 forks source link

About the pre-trained model. #3

Closed GengDavid closed 3 years ago

GengDavid commented 3 years ago

Hi @jd730, thanks for sharing your wonderful work. In the README, you said the training requires "void-suppression" pre-trained model. However, I find that in the config file, you still use the standard pre-trained model here. If I want to train your EOPSN, which pre-trained model should I use? Thx.

jd730 commented 3 years ago

Hi @GengDavid, Thank you for reporting it. 'void-suppression' pre-trained model is the correct one.

GengDavid commented 3 years ago

@jd730 Got it. Thanks for your reply. I have a further question about the 'void-suppression' pre-trained model. Does it mean pre-training the backbone model on ImageNet without the "Unknown" classes (i.e., using 1000-16=984 classes for pre-training)?

jd730 commented 3 years ago

No, it is trained on COCO without unknown class and with suppression loss for void proposals: image Please refer to the main paper for more details.

GengDavid commented 3 years ago

I see, thanks!