search

CYang0515 / NonCuboidRoom

Learning to Reconstruct 3D Non-Cuboid Room Layout from a Single RGB Image
MIT License
102 stars 13 forks source link

Wrong Predictions using custom Images #7

Open marcomiglionico94 opened 3 years ago

marcomiglionico94 commented 3 years ago

I tried to use the Structured 3D pretrained model on some custom images taken from my phone and online. I noticed that the prediction are not very correct and I wanted to know if there is any preprocessing step that needs to be done on the images. Here are some results I obtained:

0_select (2)

0_select (3)

Any help would be appreciated

bertjiazheng commented 3 years ago

Hi @marcomiglionico94,

Thanks for your interest in our work! A common failure mode of our approach (as discussed in Section 4.4 of our paper) is that the ceiling-wall and floor-wall boundaries are not precisely localized. The reason is clear since our method does not explicitly detect these two boundaries.

Best, Jia

marcomiglionico94 commented 3 years ago

Thanks @bertjiazheng any ideas of how we can visualize it in 3 dimensions like you did in the paper? Thanks, Marco

bertjiazheng commented 3 years ago

I will add this feature (same as #1) ASAP.

Best, Jia

marcomiglionico94 commented 3 years ago

Great @bertjiazheng, thanks a lot

marcomiglionico94 commented 2 years ago

I will add this feature (same as #1) ASAP.

Best, Jia

Hi any updates on this?

HardikJain02 commented 4 months ago

I will add this feature (same as #1) ASAP.

Best, Jia

Any plan to publish? If not, you can tell the approach or publish the messy code.