Reliability-Regulated Trajectory Optimization for COLMAP-free 3DGS Released on arXiv

A paper has been published on arXiv about a camera trajectory optimization method for progressive COLMAP-free 3D Gaussian Splatting. The authors propose a unified self-supervised bidirectional cyclicity mechanism that regulates trajectory estimation on two horizons: forward motion propagation with adaptive gating of kinematic warm-starts, and retrospective trajectory correction via dynamically weighted relative pose consistency constraints in a sliding window. The code has been released as open source. Read more — on the paper page.

What it means

The work directly concerns Gaussian Splatting (3DGS) — a 3D generator from Inria and MPII. Our fact sheet for this tool states that the original implementation requires 24 GB of VRAM to train to the quality claimed in the paper, and runs on CUDA via PyTorch. The new method targets the progressive scenario, where camera poses are tracked sequentially without COLMAP. The 3DGS fact sheet does not describe this mode separately: the authors of the original implementation do not specify the behavior under progressive tracking and do not describe mechanisms for correcting trajectory drift. The proposed reliability regulator is an add-on external to the base implementation that solves the problem of error accumulation in early pairwise registrations. The license for the new code is not specified by the author. The base 3DGS license restricts use to non-commercial research purposes, and this restriction must be preserved when distributing derivative works. The method is aimed at researchers who work with progressive 3DGS modes and encounter trajectory drift. For users applying the classic pipeline with COLMAP, no changes in the behavior of the base tool are claimed.