Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction

• Ultra-long video reconstruction requires causal camera and geometry estimation with bounded memory and per-frame computation despite large temporal gaps.Source: §1 · p. 1
• Local reference-frame targets keep the estimation range fixed as sequences grow, avoiding elaborate long-range cache management.The formulation supports training on short sequences while applying the model to ultra-long streams.Source: §1 · p. 2
• Global poses and scene geometry are recovered by composing adjacent-frame transformations and transforming local point maps into the global reconstruction.Source: §1 · pp. 2–3
• A temporal rotation refiner and composition-aware pose supervision reduce drift from accumulated adjacent-frame pose errors without extending temporal context.The supervision accounts for the accumulated effect of multi-step pose composition.Source: §1 · p. 3
• On Oxford Spires, ABot-Recon reduces both ATE and RPE-R by approximately 40% relative to the respective best prior results.The method achieves this across challenging long-sequence evaluations while using only local temporal context.Source: §1 · p. 3
• ABot-Recon uses a 12-frame horizon, caching KV features from the preceding 11 frames while maintaining no persistent learned long-range state.It predicts a current-camera point map, confidence map, and adjacent-frame transformation.Source: §1 · p. 3

• Feed-forward reconstruction methods replace substantial portions of classical SfM and SLAM pipelines with learned geometric priors for cameras, point maps, depth, and tracks.Source: §2.1 · p. 3
• Streaming systems propagate historical information through external spatial memories, recurrent states, cached temporal keys and values, sliding windows, or explicit spatial pointers.Source: §2.2 · p. 4
• Later methods improve long-sequence scaling through gated memories, keyframe-relative prediction, cache refresh, feature compression, adaptive state updates, and persistent attention.Source: §2.2 · p. 4
• Local-to-global approaches separate bounded local motion estimation from global trajectory recovery through composition, graph optimization, submap alignment, or loop closure.Source: §2.3 · p. 4
• ABot-Recon predicts bounded local poses in an equivariant reference frame, supervises multi-span pose chains, and refines rotations using recent context without learned long-range memory.An optional training-independent loop-closure backend can further refine trajectories at inference time.Source: §2.3 · p. 4

Local prediction formulation

• ABot-Recon predicts local point maps, confidence maps, and adjacent-frame poses, then recovers global geometry and camera motion through composition.Point maps are expressed in the current camera coordinate system, while relative poses connect adjacent frames.Source: §3.1 · p. 4, §3.2 · p. 5

Rotation refinement

• A motion–visual rotation refiner fuses recent motion and dense visual evidence, then predicts a residual rotation to correct unreliable relative-rotation estimates.A lightweight gated TCN aggregates the current pairwise feature with the K −1 preceding features while preserving causality.Source: Figure 4 · p. 6, §3.3 · pp. 6–7

Composition-aware supervision

• Composition-aware supervision compares multi-step composed transformations with ground truth instead of supervising adjacent transformations independently.Supervised frame pairs span gaps up to K −1, with weights proportional to temporal gap to emphasize accumulated chain errors.Source: §3.4 · p. 7

Training objective

• Training combines composed relative-pose losses with residual regularization and local point-map, surface-normal, and confidence losses.The residual term encourages small-magnitude and temporally consistent rotation refinements, while λ coefficients balance the supervision terms.Source: §3.4 · p. 7

• ABot-Recon achieves strong trajectory accuracy across KITTI, Oxford Spires, and VBR using compact local temporal context without persistent global memory.Source: §4.2 · p. 10
• 4.35 m average ATE is achieved on Oxford Spires without loop closure, while optional loop closure further reduces average ATE to 4.02 m.Source: §4.2 · p. 11
• ABot-Recon maintains the lowest running adjacent-frame pose error throughout KITTI-02 without progressive degradation as the processed horizon grows.Source: §4.2 · p. 11
• ABot-Recon achieves the lowest average rotational and translation errors at every evaluated KITTI multi-gap, with slower error growth at longer gaps.Source: §4.2 · p. 11
• The model processes streams at 24.45 FPS using 6.71 GB of GPU memory, while CUT3R and TTT3R use fewer resources but exhibit substantially larger long-sequence drift.Source: §4.2 · pp. 11–12
• On Oxford Spires, ABot-Recon achieves the lowest CD and highest F1 among compared methods, while remaining competitive on 7Scenes and TUM-Dynamic.Source: §4.3 · p. 12
• Composition-aware pose supervision improves consistently over the 32-frame baseline by directly constraining multi-step transformations and relative-pose consistency.Source: §4.4 · p. 13
• Increasing training clips from 32 to 128 frames reduces ATE by 19.3%, 36.7%, and 17.8% on KITTI, Oxford, and VBR, respectively.Source: §4.4 · p. 13

• The local formulation is particularly effective for long, temporally continuous streams, whereas gains are less pronounced and reconstruction is not uniformly best on compact indoor benchmarks.Source: §5 · p. 14
• Loop closure complements local prediction by adding sparse long-range constraints at inference time without changing the learned streaming formulation.Source: §5 · p. 14
• Future extensions include stronger handling of dynamic scenes, external memory, larger-scale pretraining, and selective long-range constraints when local continuity becomes ambiguous.Source: §5 · p. 14

• ABot-Recon operates entirely within a fixed twelve-frame temporal horizon.Its learned predictor avoids persistent long-range state while maintaining bounded model memory and per-frame computation.Source: §6 · p. 15
• The model predicts current-camera-frame geometry and adjacent-frame motion, then composes these local measurements into global trajectories and scene geometry.Source: §6 · p. 15
• A lightweight motion-visual rotation refiner and composition-aware pose supervision improve the reliability of multi-step pose composition.Source: §6 · p. 15
• Experiments across long driving, handheld, and landmark-scale sequences show accurate camera tracking and dense reconstruction.Source: §6 · p. 15
• The results indicate that persistent learned long-range memory is not required for effective ultra-long streaming reconstruction when local predictions remain composable.Source: §6 · p. 15

• Jiarong Han, Jingcheng Xiong, Yuzhou Liu, Ming Qian, Linzhe Shi, and Changjie Wu are listed as contributors.Source: §7 · p. 15
• Mu Xu and Ning Guo are listed as project sponsors.Source: §7 · p. 15
• Hang Zhang, Jiarong Han, and Ming Qian are listed as project leaders.Source: §7 · p. 15
• The paper expresses gratitude to eight additional supporters for their valuable support and contributions.Source: §7 · p. 15

Temporal sampling

• Video training samples are ordered sequences, with interval [a, b] specifying frame-index steps between consecutive inputs.Source: §A.1 · p. 21
• Forward fixed-stride sampling selects one interval and traverses long trajectories chronologically.Source: §A.1 · p. 21
• Adaptive forward sampling reduces the maximum interval when the remaining trajectory cannot fit the configured clip length.Source: §A.1 · p. 21
• Foldback sampling reverses traversal at sequence boundaries and resamples the stride, extending short videos while preserving local continuity.Source: §A.1 · p. 21

Multi-view construction

• Unordered multi-view sequences use pose-graph neighbors, motion and turning costs, beam search, and a minimum unique-view ratio to select valid paths.Source: §A.1 · p. 21
• US4K requires unique views, whereas HyperSim and BlendedMVS permit constrained revisiting when necessary.Source: §A.1 · p. 21

Filtering strategy

• Inspection of sampled examples and loss-spike samples identifies problematic data for dataset-specific filtering.Removing erroneous samples substantially improves model performance.Source: §A.2 · p. 21

Dataset-specific filtering

• BlendedMVS excludes 38 scenes containing sideways or upside-down images.Source: §A.2 · p. 21, Figure 8 · p. 21
• TartanAir and TartanAir-v2 invalidate erroneous sky-region depths and exclude the Ocean scene because of water-surface depth errors.Source: §A.2 · p. 21, Figure 9 · p. 22
• OmniWorld-Game removes unreliable scenes after validating cross-frame correspondences using SIFT matching and fundamental-matrix RANSAC.Source: §A.2 · p. 22, Figure 10 · p. 22
• DL3DV removes scenes with rotation above 70° or translation above 15× the scene’s median inter-frame translation, and denoises depth maps.Source: §A.2 · pp. 22–23
• HyperSim excludes two sequences with completely black RGB frames, while ScanNet++ excludes scenes whose floating mesh artifacts occlude valid depth.Source: §A.2 · p. 23
• ARKitScenes LiDAR depth maps are not used for supervision because their low resolution is insufficiently accurate.Source: §A.2 · p. 23

• 4,661 frames are processed from the full KITTI-02 sequence at stride 1.Source: B Runtime and Memory Evaluation Protocol · p. 23
• Measurements use a single NVIDIA H100 GPU and each method’s specified inference precision and input resolution.Source: B Runtime and Memory Evaluation Protocol · p. 23
• The protocol includes untimed 64-frame warm-up, streaming-state reset, GPU-preloaded inputs, and released chunk-based modes for LongStream and HorizonStream.Source: B Runtime and Memory Evaluation Protocol · p. 23

Evaluation settings

• The out-of-benchmark evaluation covers low-viewpoint robot navigation and large-scale real-world traversal without accurate ground-truth camera trajectories.These results are intended as qualitative evidence rather than quantitative comparisons.Source: C More Results · p. 23

Low-viewpoint robot navigation

• Low-mounted robot cameras create ground-dominant views with reduced visibility of distant structures, making frame-to-frame correspondence more challenging.Source: C More Results · p. 23, Figure 11 · p. 24
• Despite this domain shift, ABot-Recon produces more coherent trajectories across robot-navigation sequences than other methods.The examples provide qualitative evidence of stable trajectory estimates under less frequently represented viewpoints and motion patterns.Source: C More Results · pp. 23–24

Large-scale real-world sequences

• The four large-scale videos span urban driving, campus roads, repeated structures, intersections, and varied camera motion and visual appearance.They complement standard benchmarks with less curated and more heterogeneous long-video streams.Source: C More Results · p. 24, Figure 12 · p. 24
• ABot-Recon produces more consistent large-scale trajectory structure than compared streaming baselines across the four sequences.Its adjacent-relative-transform target remains unchanged as explored area or sequence length increases.Source: C More Results · p. 24
• Lightweight temporal correction suppresses locally correlated rotational errors before they accumulate through pose composition.Source: C More Results · p. 24