zouchuhang / LayoutNet

Torch implementation of our CVPR 18 paper: "LayoutNet: Reconstructing the 3D Room Layout from a Single RGB Image"
http://openaccess.thecvf.com/content_cvpr_2018/papers/Zou_LayoutNet_Reconstructing_the_CVPR_2018_paper.pdf
MIT License
419 stars 93 forks source link

What is the scale of the reconstructed 3D room to the actual room? #26

Open rachs opened 5 years ago

rachs commented 5 years ago

Hi ,

Thanks for sharing your work. This is very interesting and has some great use cases.

With regard to the result, is it possible to find the measurements of the actual room from the 3D reconstruction? Or in other words, how accurate is the 3d reconstruction in terms of wall width or length measurements to the actual room ?

zouchuhang commented 5 years ago

@rachs Our reconstruction (wall width/length etc.) is proportional to the camera height (which is assumed 1.0m across all images in our paper). If you want the exact reconstruction for evaluation, you need to either have a known camera height beforehand for all the panoramas or to predict the camera height from the panoramas directly.

rachs commented 5 years ago

Cool! What is the accuracy of the current model (with height assumption of 1 m) in the paper (in terms of error tolerance) with respect to the actual room measurements?

zouchuhang commented 5 years ago

@rachs You can refer to our paper (Table 1&3) for the 3D IoU evaluation comparing between our reconstruction result and the ground truth 3D layout assuming camera height 1.0m also (meaning that we have an exact camera height estimation).

vuptran commented 5 years ago

@zouchuhang, could you elaborate on how you would estimate the dimensions of the room, either from panorama or reconstruction, using known camera height and other useful camera parameters? I would like to know how LayoutNet can recover the actual dimensions of the room in terms of length x height x width of the walls, for example. Thank you.