Exporting camera parameters and poses as well as 3D points to COLMAP-style files

[yt2639]

Hii guys, has anyone figured out how to export the scene to NeRFStudio transforms.json or COLMAP-style (Cameras.txt, images.txt, points3D.txt) yet?

I have tried the solution mentioned in this issue #8. I corrected the extrinsics to be the inverse of scene.get_im_poses() as suggested in this issue. I also saved the point clouds to a .ply file by adding one line after

dust3r/dust3r/demo.py

Line 81 in 9869e71

pct = trimesh.PointCloud(pts.reshape(-1, 3), colors=col.reshape(-1, 3))

pct.export(os.path.join(outdir, 'scene.ply')).

Now I wanted to use the exported transforms.json along with the exported .ply point cloud file to train a splatfacto in NeRFStudio. However, after looking at the Viewer, the point cloud looks correct, but the camera poses are completely off.

I am wondering if anyone has successfully exported these files and verified the correctness in NeRFStudio?

Thanks!

[GGuussss]

hi, did you manage to export data to nerfstudio ?
I have the same idea. I can t get good result with colmap, but with the same dataset, i have some good result with Dust3r and Mast3r. So, i think it can replace Colmap.

[g8392]

Any progress on getting the cameras out?

[tackgeun]

Hi guys, Refer to the script scripts/run_train_infer.sh and combined codes in InstantSplat. They already implement the pipeline for exporting dust3r result into colmap-compatible format and training Gaussian splatting with processed COLMAP format.