search

nianticlabs / simplerecon

[ECCV 2022] SimpleRecon: 3D Reconstruction Without 3D Convolutions
Other
1.28k stars 120 forks source link

Using first but not last frame of a tuple as 'reference frame'? #29

Closed fengziyue closed 1 year ago

fengziyue commented 1 year ago

Hi @mohammed-amr :

Thank you for sharing this amazing work!

During reading the codes, I noticed that from here you are using the first frame instead of the last frame (stated in the paper) in each tuple as the reference frame. In this case, the model is actually using 'future frames' information for predicting the current frame depth. Is my understanding correct?

Any insights are greatly appreciated!

fengziyue commented 1 year ago

Seems in 'test_eight_view_deepvmvs' the first frame in the tuple is the last frame in actual time, but in 'train_eight_view_deepvmvs' and 'val_eight_view_deepvmvs,' the first frame is the first in actual time.

mohammed-amr commented 1 year ago

For default and dense test tuples, the first frame in the line is the reference frame. All following frames are source frames.

The sort you linked is on source frames only.

mohammed-amr commented 1 year ago

Train and val tuples use whatever combination of future and past source frames. This is fine. It's from DVMVS's tuple augmentation strategy.

Please see the tuple generation scripts for more.