Extend VGGT to dynamic reconstruction

[strongBrave]

The VGGT architecture is designed for static scene reconstruction, utilizing its paradigm and multi-loss functions to achieve robust performance in such settings. Given its effectiveness in static environments, can VGGT extend its capabilities to dynamic scene reconstruction, assuming sufficient dynamic datasets with ground truth annotations (e.g., depth, camera pose, point cloud) are available, without modifying its architecture or loss functions?

Hope you can give some insights about this.

[arctanbell]

maybe you can refer https://github.com/wzzheng/StreamVGGT