search

BaowenZ / RaDe-GS

RaDe-GS: Rasterizing Depth in Gaussian Splatting
Other
328 stars 7 forks source link

On Code and Formulas #28

Open xuncpp opened 4 days ago

xuncpp commented 4 days ago

Outstanding Work! Thank you!!

However, I didn't quite understand the CUDA code, so I have two questions to ask:

  1. About middepth in the Code What exactly does it mean? Does the author believe this value represents depth more accurately? This is because the depth_ratio in the formula for depth calculation with normal constraints is set to 0.6.

  2. Why Multiply by cos in Formula 14? Firstly, t* is the result obtained in the ray space, and the paper does not mention converting it to the camera space. Is it assumed to be approximately equal? Secondly, according to the splating method, the depth obtained in the ray space should be closer to the true value? (I am not sure) . Previous work, when calculating ( D = \sum di \alpha i Ti ), did not consider cos.

I might have some misunderstandings about certain concepts and I hope someone can answer my question, really thanks!

BaowenZ commented 4 days ago

Hi! Thank you for your attention to our work. depth and middepth represent expected depth and median depth. The third coordinate in ray space is defined as the distance( ||p|| ) from camera center (see EWA splatting Eq26). So we need to multiply it by cos to get depth. Previous works directly use center depth d, i.e. p.z. They don't calculate depth from distance, so don't consider cos.

LinzhouLi commented 3 days ago

@BaowenZ Hi! Thanks for your explaination. I wonder that if GOF also multiply this cos term? And is 0.6 a magic number that blance the expected depth and the median depth