Why forward wrapping is used in depth matching

[Metro1998]

UC-NeRF/mvs/fusion.py

Lines 68 to 71 in 1c1acc8

	# source 3D space
	# NOTE that we should use sampled source-view depth_here to project back
	xyz_src = torch.matmul(torch.inverse(intrinsics_src),
	torch.cat((xy_src, torch.ones_like(x_ref).to(x_ref.device).unsqueeze(1)), dim=1) * sampled_depth_src.reshape(batch, 1, -1))

First of all, thank you for open-sourcing such excellent work. However, I have a question: when performing geometric consistency checks, why do you reproject the source depth to the reference view? Wouldn't it be sufficient to directly perform a backward warp, projecting the reference depth to the source view and matching it directly with the source depth?