gaussian-splatting
gaussian-splatting copied to clipboard
why invDepth
Hello, I noticed that the depth regularization in both Hierarchical 3DGSand this project is implemented based on inverse depth. I would like to ask why the absolute depth of the real world is not used and what are the benefits of inverse depth compared to real depth?
I guess that they are mapping the depth (> 0.2) to a lower value for better controls? (< 50).
By the way, the world of splats is not real one, it should be called as "Camera space" instead.
Same question, and I've found that inverse depths are near-large and far-large, whereas monocular depth estimates are usually near-large and far-large. Does the loss function really work?