Thanks for your good work!
I used 40 viewpoints to shoot 360° around the human body in the center (just like the LEGO data set). Although there are not many viewpoints, I achieved good results in the simulation data set (with camera calibration ground truth), even better than Neus2. However, in experiments on real data sets with the same number and position of viewpoints, the result is very poor, far inferior to Neus2.
Is it because there is calibration noise in the real data? Does this mean that neuralangelo is more sensitive to camera noise than Neus2?
Thanks for your good work! I used 40 viewpoints to shoot 360° around the human body in the center (just like the LEGO data set). Although there are not many viewpoints, I achieved good results in the simulation data set (with camera calibration ground truth), even better than Neus2. However, in experiments on real data sets with the same number and position of viewpoints, the result is very poor, far inferior to Neus2. Is it because there is calibration noise in the real data? Does this mean that neuralangelo is more sensitive to camera noise than Neus2?