search

twke18 / HSG

Unsupervised Hierarchical Semantic Segmentation with Multiview Cosegmentation and Clustering Transformers
https://twke18.github.io/projects/hsg.html
MIT License
70 stars 6 forks source link

Do you know how to train the original Segsort to compare? #4

Closed J-JunChen closed 1 year ago

J-JunChen commented 1 year ago

Hello! Thanks for your job.

I would like to know how to train the original Segsort to compare with your method. Because the original source code of Segsort was basing on Tensorflow, did you reproduced the Segsort on Pytorch for comparsion?

twke18 commented 1 year ago

You can set every loss_type here to none.

J-JunChen commented 1 year ago

You can set every loss_type here to none.

Ok, I see. I have another question, do you know how to train and inference in VOC dataset? Because the Segsort claim that "we do not pre-train on any other segmentation dataset, such as MS COCO [44] dataset." for VOC.

twke18 commented 1 year ago

In the original paper, SegSort indeed does not train on MSCOCO. However, for fair comparison, we reproduce SegSort based on our experimental setting.

There are several difference:

  1. We don't use ImageNet-trained backbone, but train all models from scratch.
  2. We train all models on MSCOCO and test on VOC.

Hope these answer your question.

J-JunChen commented 1 year ago

Ok, Thanks.