NoteNextra-origin/content/CSE5519/CSE5519_E5.md

CSE5519 Advances in Computer Vision (Topic E: 2025: Deep Learning for Geometric Computer Vision)

VGGT: Visual Geometry Grounded Transformer

link to paper

Novelty in VGGT

Use alternating attention to encode the image.

Tip

 VGGT uses a feed-forward neural network that directly infers all key 3D attributes of a scene using alternating attention and is robust to some non-rigid deformations.

I wonder how this model adapts to different light settings for the same image, how the non-Lambertian reflectance is captured, and how this framework can be extended to recover the true color of the objects and evaluate the surface properties of the objects.