Robbyant / lingbot-map

ECCV 2026 Best Paper Award Candidate · streaming 3D reconstruction

Keep the first frames, the last few, and six tokens of everything between.

LingBot-Map is a feed-forward 3D foundation model that reconstructs scenes from a video stream. Its Geometric Context Transformer attends to three kinds of context — an anchor, a pose-reference window and a trajectory memory — so the streaming state stays compact over sequences longer than 10,000 frames.

LingBot-Map: Geometric Context Transformer for Streaming 3D Reconstruction Robbyant Team · Apache-2.0
Read the paper Get the model
  1. Anchor — the first n frames, full tokens
  2. Trajectory memory — every middle frame, six tokens each
  3. Pose-reference window — the last k frames, full tokens
  4. New frame — attends to all three
Figure 1. What a new frame attends to, drawn after Section 3 of the paper. Block height stands for retained tokens: anchor and window frames keep their image tokens; every other frame keeps six context tokens — camera, anchor and register — and drops the rest. Schematic; n, k and the frame count are illustrative.
coordinate & scale grounding

Anchor context

Monocular reconstruction is scale-ambiguous, so the first n frames attend fully to each other and fix the scale. Every later frame attends to them as fixed references.

dense local geometry

Pose-reference window

A sliding window of the k most recent frames keeps full image tokens, so each new frame registers against dense, nearby views.

long-range drift correction

Trajectory memory

Every other frame keeps six tokens with temporal position encodings — a lightweight, ordered record of the whole trajectory that helps correct accumulated drift.

A reference frame, a local window and a global map — the SLAM recipe, replaced with end-to-end learned attention.Paraphrasing the paper’s own motivation
~20 FPSat 518 × 378 input
10,000+frames in one sequence
~25,000frames in the 13-minute indoor demo

Figures as reported by the project’s README and abstract.