You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi Jingwei, thanks for the amazing code. I have a few questions about the dataset and train_affine_dorn. Would it be possible to release the depth.png? What is the args.horizontal in train_affine_dorn.py used for? Thank you.
The text was updated successfully, but these errors were encountered:
Our project is unrelated with the depth image/prediction. If you want to try depth estimation with our code, those images can be acquired from ScanNet.
Horizontal is a tricky thing. Basically, horizontal surfaces are those regions which are more "horizontal", or surfaces in which angles between the surface normal and gravity is smaller than 45 degree.
This is used for training the tangent directions. The story is that there is a natural ambiguity in the principal tangent directions (imagine a square room, there will be 4 directions at the floors). This ambiguity is harmful for training efficiency.
Fortunately, there is a clear definition of gravity vector for indoor scenes.
Therefore, we first train our network for those "non-horizontal" surfaces and estimate one of the four principal directions that closest to the gravity vector. This will give the network a good initialization. After that, we train all surfaces to predict any of the 4 directions.
it seems like args.horizontal == 1 is never used since initially args.horizontal == 0 and in the second epoch args.horizontal == 2?
By the way, how many epochs does the network need to be converged? I am training the net and it takes roughly 7 hr/epoch using single 2080ti with a batch size of 8.
Hi Jingwei, thanks for the amazing code. I have a few questions about the dataset and train_affine_dorn. Would it be possible to release the depth.png? What is the
args.horizontal
intrain_affine_dorn.py
used for? Thank you.The text was updated successfully, but these errors were encountered: