---
id: capture/lingbot-map
about:
  depiction: capture/lingbot-map.jpg
  label: LingBot-Map
  description: >-
    Open-source streaming 3D reconstruction model from the Robbyant team at Ant
    Group that estimates camera poses, depth and point clouds from a video or
    image sequence in one feed-forward pass.
  tags:
    - 3d reconstruction
    - computer vision
    - point cloud
    - video
    - open source
tool:
  form: open-source model
  group: 3D capture and reconstruction
  homepage: 'https://github.com/robbyant/lingbot-map'
  licence: Apache-2.0
  cost: free
open:
  source: open
  data: unknown
  api: unknown
  participation: unknown
  licence: Apache-2.0
provenance:
  origin: agent
  sources:
    - title: >-
        LingBot-Map: Geometric Context Transformer for Streaming 3D
        Reconstruction
      url: 'https://github.com/robbyant/lingbot-map'
    - title: >-
        GitHub - eceo-epfl/deepreefmap: Tools for reconstructing and mapping
        coral reef scenes from underwater images
      url: 'https://github.com/eceo-epfl/deepreefmap'
  created: '2026-10-17'
  reviewed: '2026-10-17'
  reviewed_by: 'Claude, from the sources listed'
  status: draft
import:
  source: openai-linkedin-scrape 2026-10-17
  id: DT043
  kind: enabling_model
  family: enabling_technology_and_adjacent_work
  priority: supporting
  verified: true
  posts:
    - P107
  posters:
    - Jonathan Sauder
domains:
  - places
---
Robbyant calls it a 3D foundation model, meaning a large model trained broadly so it can be reused for many tasks.[^1] LingBot-Map, described in the paper "Geometric Context Transformer for Streaming 3D Reconstruction" (arXiv 2604.14141), reads a video file or a folder of images and outputs camera poses, depth and a point cloud frame by frame.[^1,2] Classical structure-from-motion, which recovers 3D shape from many overlapping photos, refines its estimate by iterative optimisation; LingBot-Map does without that step.[^1,2] The repository claims about 20 frames per second at 518 by 378 pixels on sequences of more than 10,000 frames.[^1]

DeepReefMap is an EPFL toolkit for mapping coral reefs from handheld underwater video, and LingBot-Map is one of its reconstruction backends.[^2] The DeepReefMap README recommends it for long videos on a modest GPU, because memory use stays flat however long the dive is.[^2] The weights are public, with no access request.[^2]

Code is on GitHub under the Apache 2.0 licence, and the checkpoint is on Hugging Face and ModelScope.[^1] It needs an NVIDIA or AMD GPU, except on Apple Silicon, where it runs on CPU only, and it comes with a browser viewer and an offline rendering pipeline.[^1,2] The repository has updates through June 2026, and a stronger model for longer sequences has been announced.[^1]