Using MASt3R instead of DUSt3R

[anekha]

Hello
I wanted to confirm if we need to amend any files/code to use MASt3R instead of DUSt3R.
I see that in the file pred_poses_mast3r.py the poses are saved as the below:

' np.save(os.path.join(scene_path, f'dust3r_depth_{sparse_num}.npy'), depths)
confidences = np.array([confidence.detach().cpu().numpy() for confidence in confidences])
np.save(os.path.join(scene_path, f'dust3r_confidence_{sparse_num}.npy'), confidences)'

as per screenshot attached. Additionally, all the training files reference DUSt3R. So I wanted to check whether I would just run the MASt3R poses and they are saved over the DUSt3R and I run everything as DUSt3R?

That is, I run the below:

!python train_gs.py -s data/jewelry/ring_two
-m output/gs_init/ring_two
-r 1 --sparse_view_num 12 --sh_degree 9
--init_pcd_name dust3r_12
--white_background --use_dust3r

Thanks

[UranusITS]

Thank you for your question! You can proceed as described without any issues. MASt3R is fully compatible with the existing pipeline, and the poses and point clouds generated by MASt3R are more stable compared to DUSt3R.