Loop Closure Issues

[SaiSugunSegu]

Hey @Nik-V9 , Thanks for the Work.

I am using Splatam on Behavior 1k Dataset. I am facing issues at loop closure, I have read other tickets that you mention about the difficults at loop closure.

Can you suggest what i can do to improve loop closure,

• May be increase list of keyframes

[jashshah999]

I've been working on this problem from a different angle. Instead of trying to fix loop closure inside the SLAM system, I built a factor graph post-processing pipeline that takes any pose estimates (from SplaTAM, VGGT, etc.) and refines them with appearance-based loop closure.

Uses DINOv2 for finding revisited locations + ORB geometric verification + GTSAM factor graph with Cauchy robust kernel. Tested on TUM-RGBD and Replica with 70% average pose error reduction.

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

The factor graph approach could be applied as a post-processing step on SplaTAM's output to correct loop closure drift.

[jashshah999]

Loop closure in 3DGS-SLAM is fundamentally harder than in classical SLAM because you need to correct the Gaussian map alongside poses.

I built gtsam-splatfactors which addresses this by putting camera poses in an iSAM2 factor graph while keeping Gaussians optimized separately via Adam. Loop closure factors propagate corrections through the Bayes tree to all affected poses, then the Gaussians are re-optimized against the corrected trajectory.

Key design: photometric rendering residuals from gsplat expressed as GTSAM factors, so you get both the quality of differentiable rendering AND the infrastructure of factor graphs (loop closure, incremental updates, uncertainty).

https://github.com/jashshah999/gtsam-splatfactors