QUORUM — Project Page

Source for the project page of:

QUORUM: Multi-View Feature Consensus for Open-Vocabulary SLAM in Dynamic Scenes Zaid Nasser*, Mikhail Iumanov*, Tianhao Li*, Maxim Popov, Jaafar Mahmoud†, Sergey Kolyubin Biomechatronics and Energy-Efficient Robotics (BE2R) Lab, ITMO University, Saint Petersburg, Russia * Equal contribution · † Corresponding author: jaafar.a.mahmoud@itmo.ru

To be presented at the 3rd Workshop on Neural SLAM (NeuSLAM), ECCV 2026.

• Code: https://github.com/be2rlab/QUORUM
• arXiv: https://arxiv.org/abs/2604.26067

About the paper

QUORUM is an online semantic SLAM system in which multi-view high-level features vote for a dense pixel-level visual-language embedding field. That single field — extracted once per keyframe with RADSeg and PCA-compressed to D = 256 — is consumed at four stages of the SLAM stack: the optical-flow prior, the factor-graph topology, a cross-view residual inside dense bundle adjustment, and the per-pixel shape of the robust kernel. A temporal stability field aggregates cross-view embedding agreement over a keyframe's neighbourhood, separating genuinely static surfaces from movable objects and actively moving agents, and maps that to the Barron shape parameter.

The system runs on raw, uncalibrated monocular RGB video at 8–10 FPS — no intrinsics, no depth sensor, no pose prior, no list of dynamic classes.

Structure

index.html                 Single-page site (all CSS/JS inline, no build step)
static/images/             Figures, favicon, video poster
static/videos/             Hero video
static/css, static/js      Unused template leftovers (bulma, fontawesome)
QUORUM/                    LaTeX sources of the paper (not published by the site)
.nojekyll                  Serve static/ verbatim on GitHub Pages

The page has no dependencies and no build step: open index.html, or serve the directory (python3 -m http.server) and visit it. Only the Google Fonts stylesheet is fetched externally.

Figures used on the page

File on the page	Source in QUORUM/images/	Shows
static/images/overview.png	df.png	Fig. 1 — full pipeline
static/images/ark.png	ark_1.jpg	Fig. 2 — adaptive robust kernels
static/images/ablation_pca.jpg	ablation_1.jpg	Fig. 3 — PCA dimensionality ablation
static/images/grounding_replica.jpg	results.png	Fig. 4 — open-vocabulary grounding on Replica
static/images/pca_office.jpg	viz.png	Fig. 5 — in-the-wild office, phone camera
static/images/pca_replica.jpg	viz2.png	Fig. 6 — RGB-PCA colorization, office 3
static/images/stream_quad.jpg	11.png	Fig. 7 — live stream, four views
static/images/teaser_poster.jpg	frame 90 of the hero video	Poster frame for the hero video

Paper figures were downscaled to ≤1800 px wide and re-encoded as progressive JPEG for page weight.

Updating the results

Every number in the five tables on the page is transcribed from QUORUM/tables/*.tex:

Page table	LaTeX source
Table 1 — TUM-RGBD ATE	tables/slam-tum.tex
Table 2 — Replica 3D segmentation	tables/semantics-replica.tex
Table 3 — kernel ablation	tables/ablation-core.tex
Table 4 — consumer ablation	tables/ablation-consumers.tex
Table 5 — capability comparison	tables/comparison-sota-features.tex

If the paper's numbers change, update both the cells and the rank-1 / rank-2 / rank-3 badges, which mirror the \best / \sbest / \tbest highlighting in the LaTeX.

Deployment

The repository is served by GitHub Pages from the repository root; .nojekyll keeps static/ intact.

Acknowledgments

Built on the Academic Project Page Template, parts of which were adopted from Nerfies.

Website License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.