HorizonStream: Long-Horizon Attention for Streaming 3D Reconstruction

Hugging Face Daily Papers Papers
3d-reconstruction streaming transformer long-horizon-attention geometric-propagation computer-vision

Summary

HorizonStream introduces a long-horizon attention mechanism for streaming 3D reconstruction that explicitly models geometric propagation via an evidence influence kernel, achieving stable, scalable reconstruction with constant memory and linear time complexity, and generalizing to sequences over 10,000 frames.

Online 3D reconstruction requires estimating camera pose and scene geometry under strict causal and bounded-memory constraints. Existing methods often suffer from drift, jitter, or collapse on long sequences. We trace these failures to a fundamental mismatch. Streaming geometry is inherently temporally heterogeneous, with evidence ranging from short-lived correspondences to persistent global scale. However, current architectures impose uniform and pathological influence patterns. For example, sliding windows enforce hard cutoffs, while ungated recurrence and causal attention cause cache saturation and spike-like attention sinks. To resolve this, we formalize geometric propagation as an evidence influence kernel and propose HorizonStream, a long-horizon Transformer that explicitly factorizes this kernel. For the long-range temporal factor, Geometric Linear Attention learns channel-wise decay rates to enable bounded, multi-timescale propagation of geometric evidence. For the short-range spatial factor, Geometric Local Attention with Spatiotemporal RoPE performs reliable 3D matching while suppressing attention sinks. Finally, Metric Readout Tokens recover stable scale and rigid pose directly from the persistent geometric state. Extensive experiments show that HorizonStream, trained on only 48-frame clips, generalizes stably to sequences exceeding 10,000\ frames with constant memory and linear time, achieving state-of-the-art streaming 3D reconstruction performance. Project Page: https://3dagentworld.github.io/horizonstream/
Original Article
View Cached Full Text

Cached at: 05/26/26, 10:43 AM

Paper page - HorizonStream: Long-Horizon Attention for Streaming 3D Reconstruction

Source: https://huggingface.co/papers/2605.23889 Published on May 22

·

Submitted byhttps://huggingface.co/NicolasCC

ccon May 26

Authors:

,

,

,

,

,

,

,

,

,

,

Abstract

HorizonStream addresses long-term 3D reconstruction challenges by modeling geometric propagation through an evidence influence kernel, enabling stable, scalable streaming reconstruction with constant memory and linear time complexity.

Online 3D reconstruction requires estimating camera pose and scene geometry under strict causal and bounded-memory constraints. Existing methods often suffer from drift, jitter, or collapse on long sequences. We trace these failures to a fundamental mismatch. Streaming geometry is inherently temporally heterogeneous, with evidence ranging from short-lived correspondences to persistent global scale. However, current architectures impose uniform and pathological influence patterns. For example, sliding windows enforce hard cutoffs, while ungated recurrence andcausal attentioncausecache saturationand spike-likeattention sinks. To resolve this, we formalizegeometric propagationas anevidence influence kerneland proposeHorizonStream, along-horizon Transformerthat explicitly factorizes this kernel. For the long-range temporal factor,Geometric Linear Attentionlearnschannel-wise decay ratesto enable bounded, multi-timescale propagation of geometric evidence. For the short-range spatial factor,Geometric Local AttentionwithSpatiotemporal RoPEperforms reliable 3D matching while suppressingattention sinks. Finally,Metric Readout Tokensrecover stable scale and rigid pose directly from the persistent geometric state. Extensive experiments show thatHorizonStream, trained on only 48-frame clips, generalizes stably to sequences exceeding 10,000\ frames with constant memory and linear time, achieving state-of-the-artstreaming 3D reconstructionperformance. Project Page: https://3dagentworld.github.io/horizonstream/

View arXiv pageView PDFProject pageGitHub5Add to collection

Get this paper in your agent:

hf papers read 2605\.23889

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper1

#### NicolasCC/HorizonStream Depth Estimation• Updatedabout 2 hours ago

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.23889 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.23889 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Lite3R: A Model-Agnostic Framework for Efficient Feed-Forward 3D Reconstruction

Hugging Face Daily Papers

Lite3R is a model-agnostic framework that improves the efficiency of transformer-based 3D reconstruction using sparse linear attention and FP8-aware quantization. It reduces latency and memory usage by up to 2.4x while maintaining geometric accuracy on backbones like VGGT and DA3-Large.