Hey All!
I have a question that has been dogging me for a while.
I'm working on a project using extremely high resolution (5cm) Digital Surface Models (DSMs) with 5cm orthomosaics over a ~5km^2. I'm currently using the shader to calculate per-pixel normals, but get very ugly banding on the edges between quads (as the sampled pixels are out of range). This is shown below. A further issue is that the normals are calculating only on the sampled quad when at a higher quadtree depth, and aren't sampling the surrounding pixels, giving a distinct gridding phenomena.
link:[Banding](https://imgur.com/wG1rWuN)
link:[Quad Gridding](https://imgur.com/IE6bv8H)
Now, I can calculate per-pixel normals and save them to separate rbg images, but I'm trying to save on memory and already have 17 8k 16r and 17 8k BC7 RGBA images... so around 2gb in memory for images so far. So, my question: Is there is a way to sample neighbouring images efficiently and with low impact on memory and data transfer? I'd rather not load the neighbours into each shader material, but I don't know a way around it.
details:
There are 17 8192x8192 16-bit single-band heightmaps in a grid using a quadtree lod for collision with hull/domain tessellation to maximize resolution. I'm using a 4 pixel sample for each normal calculation.
↧