Some coordinates will fall outside the image area during forward splat #2
Description
Hi, thanks to your work, it is great while I have some puzzles.
During forward splat, we can compute the target coordinates according to the camera intrinsic and extrinsic. After that we can compute a target image using the source image values and the target coordinates. But the target coordinates may exceed the image area and the target image will have no values in those areas, especially in the initial stage of training. In extreme cases, the whole image will not receive any value from the original image and will not provide any gradient in loss operation.
I want to know how you deal with this problem? The paper says "To overcome this, we simply render the target frame at half the input resolution, i.e. the output image from the rendering function described in Section 3.2 is half the size of the input LDI". Why half resolution can solve this problem?