Colmap and Mast3r pose scales are inconsistent

[Grizemc]

I have a question, theoretically speaking, the pose scales of colmap and mast3R are not consistent because dust3R cannot estimate a point cloud of the same scale as colmap. Therefore, you used colmap's pose for optimization during testing. Is this correct, or was there something I thought wrong? Please correct me. Thank you

[jashshah999]

For the scale inconsistency between COLMAP and learned SfM (MASt3R/VGGT) — I built a pipeline that handles this via Sim(3) alignment across chunks with confidence weighting. Also exports directly to COLMAP format with consistent scale:

python run.py --video video.mp4 --output scene/ --export colmap

Factor graph (iSAM2 + DINOv2 loop closure) ensures global scale consistency across chunks. Works on 8GB+ GPUs.

https://github.com/jashshah999/vggt-factor-refinement

[jashshah999]

recon3d handles scale consistency via Sim(3) chunk alignment + factor graph. Outputs COLMAP format directly:

recon3d run video.mp4 -o scene/

Works on 8GB+ GPUs, 250k Gaussians in ~5 min.