How to obtain dynamic masks for Point Odyssey in DAS3R

[frigg818]

Hello, thanks for releasing DAS3R and the code.

In the paper you mention:

For successful training of this model, we produce ground-truth dynamic masks for Point Odyssey. Point Odyssey provides sparse 3D points motion information based on its 3D trajectories, from which we can generate the image-level dynamic mask by nearest interpolation to the image pixel’s 3D projection.

I could not find the corresponding script or function in the repo. Could you kindly clarify:

Where in the code this dynamic mask generation for Point Odyssey is implemented?

Or if it is not included, is there a pre-computed version available for download (similar to Sintel’s dynamic labels)?

Thanks a lot for your help!

[kai422]

Hello, thanks a lot for your interest.

The code for generating dynamic masks on Point Odyssey is not yet included in the repo. The procedure itself is fairly straightforward:

Point Odyssey provides global 3D trajectories for all points. A point can be marked as dynamic if its 3D position changes over time; otherwise it is static.

To obtain image-level dynamic masks, each pixel is first reprojected into 3D space.

Then, we use nearest interpolation to propagate the dynamic/static labels from the ground-truth 3D points to the reprojected 3D points, and this gives dynamic/static labels for pixel space.

[frigg818]

Thanks a lot for your quick reply and clarification!