The normal direction problem

[SunshineWYC]

Thank you for your excellent work! The integration of traditional MVS methods with 3DGS is remarkable.
I do have one question: when calculating the normal map, how is the direction of the normal vector determined? Should it follow the direct calculation result, or could it also go in the opposite direction?
Are there any additional considerations on this aspect?

[kcheng1021]

Hi, thanks for you recognition and good question! The direction of the normal vector is important. By taking the dot product of the vectors of camera ray(optical center to the 3D point/Gaussian) along with the normal vectors, the direction of the normal could be adjusted. Typically, the angle between the observable surface normal direction and the camera ray is greater than 90 degrees.