search

jyhjinghwang / SegSort

SegSort: Segmentation by Discriminative Sorting of Segments
https://jyhjinghwang.github.io/projects/segsort.html
MIT License
158 stars 24 forks source link

About unsupervised set-ups #1

Closed Harry-Zhi closed 5 years ago

Harry-Zhi commented 5 years ago

Hi, Very impressive work on trying unsupervised set-up for semantic segmentation.

Some questions about the codes:

  1. Does the unsupervised training still need ImageNet-pretrained weights as an initialisation?
  2. Are the results in Table. 3 also obtained with ImageNet-pretrained weights?
  3. In Fig.7, in the third "prediction" column. are the black "background" regions obtained by masking directly from Ground Truth or by indirectly voting from background patches?

Thanks.

jyhjinghwang commented 5 years ago

Hi Harry, Thanks for your thoughtful questions. 1) Yes, we do use pre-trained weights for initialization as VOC is too small as a dataset to train on from scratch. (Even with supervision, the training iterations are too many to achieve satisfactory results.) 2) Yes. In our opinion, it is a fair comparison as the supervised setting in Table 3 is also initialized the same way. 3) It is by voting from the background patches. Please let me know if you have further questions.

Jyh-Jing

Harry-Zhi commented 5 years ago

Thank you Jyh-Jing a lot for your prompt reply!

jyhjinghwang commented 5 years ago

You're welcome! :)