Query regarding use of *world_to_aligned_camera* transformation

[ShivamDuggal4]

Hi @JiamingSuen @Burningdust21

Thanks for sharing the great work!

Could you please explain the use of the transformation "world_to_aligned_camera" at the following two places for the Scannet dataset:

• https://github.com/zju3dv/NeuralRecon/blob/cd047e2356f68f60adfb923f33572d712f64b57a/datasets/transforms.py#L64

• https://github.com/zju3dv/NeuralRecon/blob/cd047e2356f68f60adfb923f33572d712f64b57a/models/neucon_network.py#L142

• https://github.com/zju3dv/NeuralRecon/blob/cd047e2356f68f60adfb923f33572d712f64b57a/models/neucon_network.py#L149

And also whats the intuition of transforming coords_batch from the world coordinate system to the middle pose of the input images ? Is the feature volume learned at the middle pose coordinate system rather than world pose ? https://github.com/zju3dv/NeuralRecon/blob/cd047e2356f68f60adfb923f33572d712f64b57a/datasets/transforms.py#L75

Thanks & Best Regards Shivam

[murrdpirate]

I'm also curious about this transformation. In my specific application, I can remove the transformation and still get good results. It doesn't seem like the choice of coordinate system should affect the CNN, but I'm not 100% sure.