Anchor3R

Anchor3R: Streaming 3D Reconstruction with Transient Anchors for Long-Horizon Visual Mapping

Paper | Code (publication pending)

Overview

Long-horizon visual mapping requires continuous camera-motion and scene-geometry estimation without allowing the model's memory to grow with the input sequence. Maintaining a coordinate system tied to the first frame or a persistent scene state can couple local geometric reasoning to accumulated errors and historical anchor bias.

Anchor3R separates local geometric measurement from global trajectory alignment. Rather than regressing every frame into a persistent global coordinate system, it uses the current frame as a temporary reference and combines three components:

• Transient Anchors and Current-Centric Prediction — predict relative poses within the active window and local geometry in the current camera's coordinate system, avoiding dependence on a permanent first-frame anchor.
• Image-Only Streaming Cache — reuse image-conditioned key/value states across windows while creating fresh pose queries for each current-frame reference, instead of carrying reference-dependent pose states forward.
• Motion Averaging and Optional Loop Reinsertion — combine overlapping relative-pose measurements into online trajectories; offline optimization can incorporate reinserted loop frames to reduce drift and align local geometry into a global reconstruction.

Trained on 48-frame clips, Anchor3R generalizes to sequences exceeding 10,000 frames. With a fixed image resolution and active window size, its streaming model uses bounded working memory and linear total forward computation in sequence length. It supports long-horizon visual mapping from ordered RGB images or videos, producing camera poses, focal lengths, depth maps and point clouds.

These memory and computation properties describe the streaming model, not the entire export and optimization pipeline. The current CLI retains predictions, constructs trajectories and saves results at sequence end; stored outputs and offline processing are not a constant-total-memory guarantee.

Quick Start

This model repository contains Anchor3R.pt, this model card, the weight license in LICENSE, SHA256SUMS, and weight usage instructions. Inference source, configurations, evaluation tools and tests belong to the separate code repository; they are not duplicated here.

Install the matching Anchor3R code release and its documented dependencies, and place the downloaded Anchor3R.pt at checkpoints/Anchor3R.pt. Then run:

python infer.py \
  --images /path/to/images \
  --checkpoint checkpoints/Anchor3R.pt \
  --output outputs_anchor3r/demo \
  --save-geometry

For video input, replace --images /path/to/images with --video /path/to/input.mp4. Saved camera poses use the world-to-camera (w2c) convention. Do not assume that raw translations or depth are in meters.

The first release contains inference code and tensor-only weights, not training code or optimizer states. Optional loop closure requires separate SALAD and DINOv2 weights and explicit --infer-mode offline --loop. Use ordered RGB frames; calibrated preprocessing expects rectified images. This is a research model, not a safety-critical localization system.

Release Variant and Model Sources

The maintainer confirms that these weights match the project-page demos. They are a release variant with results that differ slightly from those reported in the paper, not a claim to reproduce every paper result exactly.

The maintainer confirms that initialization and training-data sources follow the Anchor3R paper: VGGT initialization and a mixture of real and synthetic datasets described in its training-data section. See the paper for the method and source description; dataset revisions and exact split/overlap details are not specified by this model card. Training code is not included in this release.

Citation

@article{tao2026anchor3r,
  title={Anchor3R: Streaming 3D Reconstruction with Transient Anchors for Long-Horizon Visual Mapping},
  author={Tao, Peilin and Cheng, Chong and Du, Yuansen and Song, Caiwei and Chen, Zhengqing and Guo, Xiaoyang and Yin, Wei and Ren, Weiqiang and Zhang, Qian and Cui, Hainan and Shen, Shuhan},
  journal={arXiv preprint arXiv:2606.05035},
  year={2026}
}

License

The Anchor3R model weights are licensed under Apache-2.0. The model distribution includes the full license text in LICENSE.

Contact

For usage questions and project feedback, contact Peilin Tao at taopeilin2023@ia.ac.cn. Use the same address for private security reports; do not post vulnerability details or credentials in public issues.

Release Status

This model repository is prepared as polar-explorer/Anchor3R on Hugging Face. After upload, record the immutable model revision in the code checkout's checkpoints/release.json; it is a download pin maintained with the source, not a file required in this model repository.

The selected code repository is polar-explorer/Anchor3R on GitHub; its creation and public availability have not yet been verified. The project page will use GitHub Pages at https://polar-explorer.github.io/Anchor3R/ (deployment pending). The matching source revision remains to be recorded.