Closed wing158 closed 4 months ago
720*800,
Hello, here are some of my thoughts on your data:
I believe the problem occurred when you ran COLMAP. Your images actually have a fixed camera pose and white backgrounds, so the camera parameters cannot be estimated by multi-view stereo.
For this type of data, it lacks multi-view information. I hold that dynamic NeRF or dynamic Gaussians cannot be trained, which heavily rely on multi-view photometric consistency.
I strongly suggest you to try those approaches utilizing human body priors, such as Animatable Gaussians, GaussianAvatar, and 3DGS-Avatar. Using SMPL model to represent human shapes and serve as a basis for dynamic deformation makes it possible to reconstruct such scenes.
But I believe these approaches are still difficult on your data, since it seems to be generated by some diffusion model (e.g. AnimateAnyone) and temporal consistency is not promised.
已解决,用其他扣图软件完成绿幕图片保存并处理,可以完成colmap。 总结:对视频一定要求进行筛选,实现部分角度输出。