LongStream：长序列流式自回归视觉几何
LongStream: Long-Sequence Streaming Autoregressive Visual Geometry
提出规范解耦的流式视觉几何模型，通过预测关键帧相对位姿和正交尺度学习，将长序列外推转化为恒定难度的局部任务，并利用缓存一致性训练与周期性刷新抑制注意力偏差，为在线长序列三维重建提供可复用方案。
原文依据 We introduce LongStream, a novel gauge-decoupled streaming visual geometry model for metric-scale scene reconstruction across thousands of frames under a strictly online, future-invisible setting.
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论长序列流式三维重建仍是一个重要的开放挑战。现有自回归模型在处理长序列时常常失败，因为它们将位姿锚定到第一帧，导致注意力衰减、尺度漂移和外推误差。我们提出 LongStream，一种新颖的规范解耦流式视觉几何模型，用于在严格在线、未来不可见设定下对数千帧进行度量尺度的场景重建。我们的方法包含三方面。首先，我们丢弃第一帧锚点并预测关键帧相对位姿。这将长程外推重新表述为恒定难度的局部任务。其次，我们引入正交尺度学习。该方法将几何与尺度估计完全解耦以抑制漂移。最后，我们识别出 Transformer 中的注意力偏差问题，包括注意力汇依赖和长期 KV 缓存饱和。我们提出缓存一致性训练结合周期性缓存刷新。该方法在超长序列上抑制注意力偏差和污染，并缩小训练与推理之间的差距。实验表明，LongStream 达到最优性能，能够在公里级序列上以 18 FPS 实现稳定、度量尺度的重建。
英文原文摘要 1279 字符
Long-sequence streaming 3D reconstruction remains a significant open challenge. Existing autoregressive models often fail when processing long sequences because they anchor poses to the first frame, leading to attention decay, scale drift, and extrapolation errors. We introduce LongStream, a novel gauge-decoupled streaming visual geometry model for metric-scale scene reconstruction across thousands of frames under a strictly online, future-invisible setting. Our approach is threefold. First, we discard the first-frame anchor and predict keyframe-relative poses. This reformulates long-range extrapolation into a constant-difficulty local task. Second, we introduce orthogonal scale learning. This method fully disentangles geometry from scale estimation to suppress drift. Finally, we identify attention bias issues in Transformers, including attention-sink reliance and long-term KV-cache saturation. We propose cache-consistent training combined with periodic cache refresh. This approach suppresses attention biases and contamination over ultra-long sequences and reduces the gap between training and inference. Experiments show that LongStream achieves state-of-the-art performance, enabling stable, metric-scale reconstruction over kilometer-scale sequences at 18 FPS.