AIAny
Icon for item

Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction

Performs causal, bounded‑memory streaming 3D reconstruction by caching KV features from only the preceding 11 frames, predicting a per‑frame point map and adjacent relative pose, and composing these local predictions into a global trajectory; includes a lightweight rotation refiner and composition‑aware loss to limit drift.

Introduction

Long video streams reveal a mismatch between learned temporal state and the geometric evidence needed for stable reconstruction: some cues are fleeting while others (scale, global rigidity) persist over arbitrarily long horizons. The core insight behind this work is counterintuitive but simple — you do not need persistent learned long‑range memory to achieve stable, metric‑scale streaming reconstruction. By designing predictions whose targets are reference‑frame equivariant and by recovering global pose and geometry via sequential composition, a strictly local learned state can generalize stably to very long sequences while keeping per‑frame memory and compute bounded.

Key Findings
  • Local‑only state with equivariant targets: the model caches KV features from only the previous 11 frames, predicts a point map in the current camera frame plus an adjacent relative pose, and composes poses to recover a global trajectory. So what: training and inference complexity remain independent of total sequence length, avoiding cache saturation and unbounded memory.
  • Drift mitigation by light refinement and loss design: a compact motion‑visual rotation refiner and a composition‑aware pose loss reduce accumulated rotation and scale drift when composing many local relative poses. So what: the method reduces long‑horizon error compared to prior streaming models that rely on persistent memory or large context windows.
  • Empirical long‑horizon stability: evaluated on long‑sequence benchmarks (e.g., Oxford Spires) the approach yields substantially lower absolute and relative pose errors versus prior streaming methods while remaining efficient to run on a single GPU. So what: practitioners can process kilometer-scale streams with constant per‑frame resources and limited model state.
Who it fits and trade-offs

Great fit if you need causal, bounded‑memory 3D reconstruction from long camera streams (e.g., driving sequences or extended walk‑throughs) and prefer a simple, composition‑based global recovery pipeline without loop‑closure dependencies. Look elsewhere if your application demands immediate globally consistent dense maps at revisit time (you may want methods with explicit loop closure or large persistent memories), or if you require per‑frame outputs tied to a fixed global reference without sequential composition.

Where it sits relative to alternatives

Compared with approaches that learn persistent compressed memories or multi‑scale trajectory tokens, this work deliberately avoids learned long‑range state and instead leans on geometric equivariance and composition. That simplifies state management and training while still delivering strong long‑horizon robustness in practice.

Information

  • Websitearxiv.org
  • OrganizationsAlibaba Group
  • AuthorsJiarong Han, Jincheng Xiong, Yuzhou Liu, Linzhe Shi, Changjie Wu, Ning Guo, Mu Xu, Hang Zhang, Ming Qian
  • Published date2026/08/27

More Items

Proposes VLAct, a representation-centric continued pre-training method for Vision-Language-Action models that preserves VLM priors and enforces cross-embodiment action semantics to turn limited robot trajectories into transferable visual-action representations; shows strong gains and sample efficiency on multiple VLA benchmarks using modest compute.

Rewrites physical scenes as executable world programs (e.g., MuJoCo scene descriptions) and uses an agentic abductive loop to propose, execute, render, verify, and iteratively refine those programs from videos or text. Verified executable worlds supply scalable physical supervision for training vision–language models.

Turns a flow-matching image generator's self-exploration into dense, per-step supervision without a pretrained teacher; it branches the student's next-state into stochastic SDE candidates, scores them against a deterministic self-reference, and applies an advantage-weighted pull–push velocity regression with reward-level fusion for multi-objective alignment.