search

mseg-dataset / mseg-api

An Official Repo of CVPR '20 "MSeg: A Composite Dataset for Multi-Domain Segmentation"
Other
249 stars 45 forks source link

Where is supplement mentioned in the paper #19

Closed steven5401 closed 4 years ago

steven5401 commented 4 years ago

Hello, I want to know the datasets that were not used in MSeg and the reasons for not including them. However, I can not find supplement through google. Can you give me the link of supplement?

johnwlambert commented 4 years ago

Hi @steven5401 , thanks for your interest in our work. The supplement will be included in an upcoming expanded Arxiv preprint. Here's a table that should address your question:

Screen Shot 2020-09-01 at 10 07 37 PM

Let us know if you have further questions.

steven5401 commented 4 years ago

Thank you very much. Why didn't you choose scannet as training set, but test set?

johnwlambert commented 4 years ago

No problem. We had a large number of concerns about the quality of the 2d labels in ScanNet. As you know, ScanNet is also annotated in 3d, not 2d. Unlike Matterport3d and Stanford-2d-3d, ScanNet does have a 2d label benchmark. Because of its noisy nature, we didn't want to train on it, but it is a good proxy for performance in the real-world.

Here are a few examples from ScanNet:

Book vs. bookshelf seems arbitrary:

Screen Shot 2020-09-02 at 11 18 57 AM

The non-white pixels in the center should be "Wall", but include "blinds"

Screen Shot 2020-09-02 at 11 18 30 AM

The door label extends onto the glass window:

Screen Shot 2020-09-02 at 11 18 09 AM

Bed and backpack boundaries are poorly aligned:

Screen Shot 2020-09-02 at 11 16 43 AM

Bicycle extent is quite incorrect:

Screen Shot 2020-09-02 at 11 16 26 AM
steven5401 commented 4 years ago

I understand it. Thank you for the prompt reply.