Weijing website

Computer vision grad class at UCSB ECE 281B project 1 where we went though pipeline of making 3D construction of Kavli building given many pictures of it from different views.

Step 1

Made manual keypoint matches by hand and computed homography and fundemental matrix.

Step 2

Use superpoint+superglue to automate keypoint selection

Step 3

Use colmap to make keypoint matches to make 3d point cloud.

Step 4

Use gaussian splatting to render point cloud.

Thoughts

Could have got better reconstruction if had more views (get entire building in one shot like aerial view) and also if utilized GPS metadata, or used depth map like depth pro.