CrossFit-GS uses out-of-fold cross-validation to fix pose drift in COLMAP-free 3D Gaussian Splatting
kwangmoo_yi · x · 2026-10-22
The paper CrossFit-GS tackles a core failure mode of pose-free 3D Gaussian Splatting: jointly optimizing camera poses and scene representation entangles the two, so errors reinforce themselves and trajectory drift accumulates.
- Method: split the input sequence into interleaved subsets and maintain two complementary submaps. Each candidate pose is cross-refined against a frozen opposite submap that never saw that frame, providing an independent geometric reference.
- Checkerboard validation and spatial-coverage safeguards reject unreliable updates; only verified poses are committed. After refinement, intermediate submaps are discarded and the final global representation is retrained from scratch with fixed poses.
- Results: across multiple benchmarks, CrossFit-GS substantially reduces cumulative drift and consistently improves both camera pose estimation and novel-view synthesis over prior methods.
Related event: CrossFit-GS: Out-of-Fold Pose Estimation Reduces Drift in COLMAP-Free 3DGS(2 posts)→
More from Multimodal
- Image Edit Showdown: Qwen 2.1 Beats Fibo 2 and Klein 9b, but Fibo Breaks Above 2MP — Technical_Fish_9638 · 2026-10-22
- Combining CozyClay 3D keyframes with Maestro to generate full videos — cocktailpeanut · 2026-10-22
- AI-generated VHS-style demon queen video showcases striking cinematic quality — brythehuman · 2026-10-22
- Moss image edit methods compared: a 6,000-step LoRA makes all the difference — Sixhaunt · 2026-10-22
- Hermes agent now edits videos directly in DaVinci Resolve from a script — alexcovo_eth · 2026-10-22
- Thrixel adds auto-rigging for characters, creatures, and robots with Mixamo/Unity/Roblox support — RanaHanocka · 2026-10-22