search

YuelangX / Gaussian-Head-Avatar

[CVPR 2024] Official repository for "Gaussian Head Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians"
Other
703 stars 40 forks source link

Why padding the sides of training images? #27

Open NikoBele1 opened 1 month ago

NikoBele1 commented 1 month ago

Hi, in your preprocessing code you add a 200px width black padding to both sides of the training images and downscale, instead of simply cropping 2048 x 2048 out of the original Nersemble images which are 2200 x 3208. Why is that?

YuelangX commented 1 month ago

@NikoBele1 Actually, I cropped 2600x2600 and resize it to 2048x2048 to ensure more content is included in the image.