Sparse Anchors for Visual Multi-Session Registration in Changing Environments

Yunus Talha Erzurumlu, Niklas Trekel, Saurabh Gupta, Nico Koltermann, Olga Vysotska, Jens Behley, Louis Wiesmann, Cyrill StachnissarXiv 2026 · arXiv

TL;DR

This paper introduces a robust visual registration pipeline that aligns new robot trajectories to old maps by identifying a few "sparse anchors"—highly reliable, geometrically verified pose constraints. By combining these anchors with the query session's internal relative motion and refining them against a Gaussian feature field, the system achieves near-perfect registration in environments undergoing massive structural and appearance changes.

What Problem Does This Paper Address?

Imagine a robot tasked with monitoring a construction site over several months. In May, it maps a floor of bare concrete and open studs. By December, walls are up, finishes are applied, and the lighting has completely changed. Standard visual localization often fails here for two reasons:

  1. Structural Change: The "landmarks" the robot saw in May simply don't exist anymore.
  2. Perceptual Aliasing: Construction sites are full of repetitive structures (identical hallways, rows of pillars) that lead to "false positive" matches, where the robot thinks it's in one room but is actually in another.

Existing methods like Hierarchical Localization (HLoc) try to localize every frame independently, which leads to catastrophic failures when matches are ambiguous. Direct 3D map registration also struggles because the geometry itself has evolved too much for traditional point-cloud alignment to work reliably.

Methods

The authors propose a five-stage pipeline designed to favor precision over density. Instead of trying to localize every image, they look for a small set of "anchors" they are absolutely sure about.

Fig. 2: Overview of our visual multi-session registration pipeline. We first select distinctive cross-session retrievals and establish local feature correspondences between query and reference observations. Generalized PnP with RANSAC uses the reference-map geometry to reject correspondences that do not support a geometrically consistent multi-camera pose, producing locally verified absolute pose hypotheses. Since a locally plausible pose may still correspond to an incorrect location, sequence consensus evaluates the hypotheses jointly with the within-session query trajectory and retains only spatially distributed hypotheses that support a common query-to-reference similarity alignment. The resulting sparse absolute anchors are refined by featuremetric alignment against the reference Gaussian feature field and finally combined with within-session relative-pose constraints and consensus-aligned priors in a pose-graph optimization to recover the complete query trajectory.
  1. Reliable Retrieval: They use a "spatial far-margin" criterion for place recognition. Instead of just checking if the best match is much better than the second-best, they check if it's better than any match outside a certain radius. This filters out ambiguous matches in repetitive environments.
  2. Geometric Verification: They use Generalized PnP with RANSAC to estimate a multi-camera body pose. This ensures that the local features matched between the query and the reference map are geometrically consistent.
  3. Sequence Consensus: This is a crucial "sanity check." They use the query trajectory's internal consistency to see if the proposed anchors all agree on a single global alignment (a similarity transformation). Outliers that don't fit the group are discarded.
  4. Featuremetric Refinement: To get sub-centimeter accuracy, they refine the anchors by aligning query features directly against a Gaussian feature field—a 3D Gaussian Splatting representation where each Gaussian carries a learned feature vector (DINOv3).
  5. Pose-Graph Optimization (PGO): Finally, they tie everything together in a graph where the refined anchors act as "pins," the within-session relative poses act as "springs," and consensus-aligned priors keep the whole trajectory from drifting.

Datasets

The approach is evaluated on the HILTI26 construction-site dataset.

  • Why it matters: Unlike many SLAM datasets that span hours or days, HILTI26 covers an active construction project over eight months.
  • Challenge: It captures massive structural changes (e.g., rooms being built) and difficult-to-match textures, providing 15 challenging cross-session experiments.

Key Insights and Ideas

  • The "Spatial Far-Margin" is a Game Changer: In repetitive buildings, the top-2 matches are often just adjacent frames of the same place. By comparing the best match to the best spatially distant match, the authors effectively ignore local ambiguity and focus on global distinctiveness.
Fig. 3: Precision–recall trade-off of VPR confidence criteria. Re- trievals are considered correct if the retrieved reference observation lies within 2 m of the ground-truth position. Our spatial far-margin achieves substantially higher precision across a broad range of recall than the top-1 retrieval score and the top-1–top-2 margin.
  • Gaussian Splatting for Features: While most use Gaussian Splatting for photorealistic rendering, this work uses it as a differentiable "feature field." This allows them to perform direct feature-to-map alignment, which is more robust to lighting changes than raw pixel intensity.
  • Sparse is Sufficient: You don't need a match for every meter of travel. If you have a few rock-solid anchors and a decent internal estimate of your own motion (from SfM or Odometry), you can "bridge" the gaps where the environment has changed too much to recognize.

Key Results

The proposed method significantly outperforms state-of-the-art baselines in both accuracy and reliability:

  • Recall: Achieves 94.4% recall within 1m and 100% within 5m. For comparison, HLoc (a standard baseline) peaks at only 65.3% recall within 1m.
  • Accuracy: The translation error (RMSE) is reduced to 0.60m, nearly 40% better than the next best method.
  • Qualitative Alignment: Featuremetric refinement visibly corrects small misalignments that geometric verification alone misses, ensuring the query images line up perfectly with the reference map's structures.
Qualitative effect of featuremetric pose refinement correcting residual misalignment.

Limitations and Future Work

  • Scale Ambiguity: The method currently relies on sequence consensus to estimate the metric scale of the query trajectory. If the trajectory is too short or lacks enough anchors, scale estimation might fail.
  • Initial Trajectory Quality: The approach assumes the "within-session" relative geometry (the "query trajectory") is locally consistent. If the underlying SLAM/SfM system used to generate the query session has massive drifts or loops, the PGO might struggle.
  • Rejection vs. Failure: If no stable consensus is found, the system rejects the session rather than providing a wrong answer. While this is great for safety, future work could explore how to recover from "partial" consensus in extremely fragmented environments.

Summary updated

Abstract

Repeated robotic monitoring requires localization across visits to the same environment, even if the environment changes. Such changes can substantially reduce persistent visual content, and repetitive structures can lead to incorrect cross-session correspondences. In this paper, we address visual multi-session registration of a new data recording (the query session) to a previously reconstructed reference map. Under substantial visual and structural changes, cross-session localization may be possible only for parts of the trajectory. In our approach, we therefore rely on a sparse set of reliable absolute pose constraints and combine them with within-session relative-pose information to register the complete trajectory to the reference map. We first estimate absolute alignment hypotheses through visual place recognition, local feature matching, and geometric verification. We then use sequence consensus to reject hypotheses that are inconsistent with a common query-to-reference alignment and retain the remaining poses as absolute anchors. These anchors are refined through featuremetric alignment against a Gaussian feature field and combined with within-session relative-pose constraints in a robust pose-graph optimization. We evaluate our approach on 15 cross-session experiments from the challenging HILTI26 construction-site dataset, covering substantial scene changes and difficult-to-match environments. Our method achieves 94.4% recall within 1m and 100.0% within 5m, outperforming independent image localization, which peaks at 65.3% and 70.7%, respectively, as well as direct registration of independently reconstructed Structure-from-Motion geometry. Our results show that sparse, reliable absolute anchors can provide accurate and complete trajectory registration even when reliable cross-session correspondences are available only for part of the query session.
BibTeX
@misc{erzurumlu2026sparse,
  author = {Yunus Talha Erzurumlu and Niklas Trekel and Saurabh Gupta and Nico Koltermann and Olga Vysotska and Jens Behley and Louis Wiesmann and Cyrill Stachniss},
  title = {{Sparse Anchors for Visual Multi-Session Registration in Changing Environments}},
  year = {2026},
  eprint = {2611.05084},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.05084},
}
Figure