How to get the pointcloud reconstruction

[charming2992]

Thank you very much for your great work and open-sourcing this project!
May i ask how to get the pointcloud reconstruction. The demo just show query points tracking?

[Lijiaxin0111]

Thanks for your interest.

The demo package contains both sparse query tracking and dense point-cloud reconstruction.
More concretely, the reconstruction comes from these functions:

1. vis/build_like_demo_for_worldtrack.py

• build_worldtrack_demo_package(...)
• around the call to _build_uv_grid(...)
• and then the call to _export_demo_data(...)

In this function, we first build a dense image grid:

• point_query_uv_px = _build_uv_grid(...)
  Then we pass these dense query points into:
• _export_demo_data(...)

2. vis/build_like_demo.py

• _export_demo_data(...)
  Inside _export_demo_data(...), the actual dense 3D reconstruction is produced here:
• points_xyz_ref0, points_vis, points_conf, _ = _infer_point_cloud_ref0(...)
  This is the core point-cloud reconstruction step.

The reconstructed outputs are then stored in:

• point_xyz_ref0 point_visibility point_confidence point_rgb
  and finally written into the demo package JSON as:
• points.xyzRef0 points.visibility points.confidence points.rgb`