zhengzhang01 / Pixel-GS

[ECCV 2024] Pixel-GS Density Control with Pixel-aware Gradient for 3D Gaussian Splatting
Other
82 stars 7 forks source link

about depth #7

Open MELANCHOLY828 opened 2 weeks ago

MELANCHOLY828 commented 2 weeks ago

Hello, I have tried to optimize the method provided in the paper for my application. However, I found that the scaling_factor is either 1 or 0, indicating that the depth is either very large or zero. I would like to ask if this is normal, as it seems that this removal of floaters has no effect.

zhengzhang01 commented 2 weeks ago

Yes, this is normal. For a given view, the scaling_factor value of the Gaussian outside the view frustum is 0. Inside the view frustum, points farther from the camera do not scale the gradient, and the scaling_factor value is 1. As shown in the diagram, the red area represents the visible region, and only a small portion of the blue region undergoes gradient scaling. 2

Yes, this is normal. For a given view, the scaling_factor value of the Gaussian outside the view frustum is 0. Inside the view frustum, points farther from the camera do not scale the gradient, and the scaling_factor value is 1. As shown in the diagram, the red area represents the visible region, and only a small portion of the blue region undergoes gradient scaling.

If you still notice floaters, you can adjust the depth_threshold parameter. The intuition behind this method is to suppress the splitting of Gaussians near the camera by reducing their gradients, thereby suppressing the floaters.