PaperlayerView exact revision on arXiv

Source-linked AI summary

Zebra3D: Efficient Hybrid Foundation Models For Scalable 3D Reconstruction

Utkarsh Saxena, Ruijing Yang, Mehdi Rezagholizadeh, Akash Haridas, Vikram Appia, Emad Barsoum

arXiv:2610.12727v1cs.CVcs.LG

TL;DR

Feed-forward 3D foundation models face a quadratic global-attention cost as input view sets grow, and uniform attention replacement does not account for stage-wise conversion tolerance. Zebra3D allocates linear-centered and tile-sparse attention across pretrained stages, yielding strong accuracy–efficiency trade-offs across three backbones, including refined VGGT results close to dense fine-tuning.

  • Problem

    Dense global attention scales quadratically with input views, and uniform sparse or linear replacement does not account for pretrained stages’ differing tolerance to approximation.

    Source: §1 · pp. 1–2

  • Method

    Zebra3D assigns linear-centered attention to conversion-tolerant stages and tile-sparse attention with landmark summaries to sensitive stages, using closed-form fitting for training-free conversion.

    Source: §1 · pp. 2–3, §3 · p. 3

  • Results

    80.5% RealEstate10K pose AUC@30 at 94% mean attention sparsity is within 1.1 percentage points of dense VGGT fine-tuned on the same data and schedule.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Closed-form conversion allows immediate use without gradient updates, while optional refinement combines task supervision with output and relational feature distillation.

    Source: §1 · pp. 2–3

    The work uses existing pretrained models and standard research datasets, with no human-subject experiments or new personal-data collection; deployments should respect consent, dataset licenses, privacy, and applicable law.

    Source: Ethics Statement · pp. 10–11

Abstract

from arXiv · show

Feed-forward 3D foundation models reconstruct scenes from unposed images, but the quadratic cost of global attention limits their scalability to large view sets. We present Zebra3D, a layerwise sparse-linear hybrid attention framework that improves the efficiency of pretrained 3D foundation models without training from scratch. Across VGGT, HunyuanWorld-Mirror, and $\pi^3$, we observe a broadly consistent depth-wise pattern: early stages are relatively tolerant to linear replacement, while sensitive middle-to-late stages tend to exhibit more concentrated attention. Zebra3D exploits this structure by assigning linear-centered attention to tolerant stages and tile-sparse attention to sensitive stages, using landmark summaries to preserve unselected context. Closed-form affine fitting enables training-free conversion, while optional post-conversion training combines task supervision with output and relational feature distillation. Across all three backbones, Zebra3D achieves strong reconstruction and camera-pose accuracy at high attention sparsity, with training-free variants remaining close to their dense counterparts. End-to-end speedups increase with input view count, reaching up to $7.1\times$ at 1,024 views. After post-conversion training, Zebra3D-VGGT reaches 80.5\% RealEstate10K AUC@30 at 94\% mean attention sparsity, within 1.1 percentage points of dense VGGT fine-tuned on the same data and schedule.

1 Introduction

Feed-forward 3D foundation models predict camera poses and dense geometry from unposed views, but dense global attention remains a quadratic bottleneck. Zebra3D uses observed stage-wise differences in linearization tolerance to allocate hybrid attention, with reported accuracy–efficiency trade-offs across three backbones.

1 Introduction

  • Feed-forward 3D foundation models jointly predict camera poses and dense scene geometry from multiple unposed views in one forward pass, without per-scene optimization.The cited models include VGGT, HunyuanWorld-Mirror, and π3.Source: §1 · p. 1
  • Global attention takes 88% of VGGT’s end-to-end runtime at 512 views, while uniform sparse or linear replacement overlooks stage-specific tolerance to approximation.Global attention scales quadratically with the number of input views.Source: §1 · pp. 1–2
  • Across VGGT, HunyuanWorld-Mirror, and π3, middle stages are especially sensitive to linear replacement yet have concentrated attention, while preceding and following stages tolerate linear approximation.The authors use this alignment to motivate sparse attention for concentrated interactions and linear attention elsewhere.Source: §1 · pp. 2–3
  • Zebra3D replaces dense global attention with linear-centered attention in tolerant stages and tile-sparse attention in sensitive stages, using landmarks to summarize omitted tiles.Closed-form affine fitting enables training-free conversion; optional post-conversion refinement combines task supervision with output and relational feature distillation.Source: §1 · pp. 2–3
  • 80.5% RealEstate10K pose AUC@30 at 94% mean attention sparsity is within 1.1 percentage points of dense VGGT fine-tuned on the same data and schedule.At 79% mean attention sparsity, training-free variants remain within 0.4–0.7 percentage points of dense RealEstate10K pose AUC@30; at 1,024 views, speedups are 6.2× for VGGT, 5.5× for HYWM, and 7.1× for π3.Source: §1 · p. 2

2 Key Related Works

Prior work reduces attention cost through token and context compression or more efficient attention operators. Zebra3D instead measures conversion sensitivity to assign linear and tile-sparse attention across stages of a pretrained model.

2 Key Related Works

  • Token-compression methods merge image tokens using saliency, geometry, temporal, or confidence cues, while other approaches subsample key/value context or compress views into descriptors.Examples include FastVGGT, LiteVGGT, HTTM, Co-Me, AVGGT, FlashVGGT, and RegimeVGGT.Source: §2 · p. 3
  • Efficient-attention approaches include selected-block evaluation, distilled sparse linear attention, and compressed context with selected full-resolution interactions.Zebra3D differs by using measured conversion sensitivity to assign linear and tile-sparse attention to different pretrained-model stages.Source: §2 · p. 3
  • VGG-T3 adapts a fixed-size scene MLP at test time, rather than assigning attention operators across pretrained-model stages.The passage distinguishes this test-time adaptation approach from Zebra3D’s stage-wise attention allocation.Source: §2 · p. 3

3 Method

Zebra3D replaces pretrained global-attention stages with a hybrid: linear-centered attention in stages tolerant to linearization and tile-sparse attention in sensitive stages. It selects stages by measured sensitivity and optionally refines the converted model with task supervision and distillation.

3 Method

  • Zebra3D assigns linear-centered attention to conversion-tolerant stages and tile-sparse attention to sensitive stages, replacing the dense global-attention stack.Stage sensitivity is measured by replacing one global-attention stage at a time and comparing output deviations with the dense teacher.Source: §3 · p. 3, §3.1 · p. 3

3.1 Sensitivity-Guided Stage Allocation

  • On VGGT, output deviations from linear replacement peak in stages 11–15, while sensitive stages retain more attention mass in the top 20% of key/value blocks.This concentration-versus-tolerance pattern also extends to HYWM and π3; stages with the lowest sensitivity scores are assigned to linear attention for a target linear fraction.Source: §3.1 · p. 3

3.2 Linear-Centered Attention with Closed-Form Initialization

  • Linear-centered attention reuses pretrained projections and fits an affine correction around the value mean to adjust linear outputs toward dense softmax outputs.Centering anchors the approximation at value-mean pooling; the correction is fitted from frozen-teacher activations without gradient updates to inherited weights.Source: §3.2 · pp. 4–5

3.3 Tile-Sparse Attention with a Landmark Tail

  • Tile-sparse attention selects spatial key tiles for token-level softmax interactions and summarizes unselected tiles with mean-key and mean-value landmarks.Selected tokens and landmarks share one softmax normalization, preserving detailed interactions in selected regions and compressed context elsewhere.Source: §3.3 · p. 5

3.4 Lightweight Refinement via Knowledge Distillation

  • Optional refinement combines task supervision with cosine, relational, intermediate-feature, and output distillation from a frozen dense teacher.Relational distillation samples cross-view token pairs, and refinement adds no inference-time computation.Source: §3.4 · pp. 5–6

4 Experiments

Experiments evaluate released VGGT, HYWM, and π3 checkpoints on reconstruction, camera-pose, and long-sequence benchmarks, comparing Zebra3D with dense and accelerated baselines. Results cover training-free conversion, post-conversion training, efficiency, and ablations of stage placement, attention operators, and training objectives.

4.1 Experimental Setup

  • Experiments use released VGGT, HYWM, and π3 checkpoints; evaluation covers point-map reconstruction, camera pose, long sequences, and wide-baseline reconstruction.Training-free conversion fits projections on 13 held-out calibration scenes without gradient updates; Zebra3D-KD refines all student parameters for 3,000 updates using a frozen dense teacher.Source: §4.1 · p. 6

4.2 Main Results

  • At approximately 79% mean attention sparsity, training-free 0.5L/0.6S retains Re10K AUC-30 within 0.007, 0.004, and 0.006 of dense VGGT, HYWM, and π3, respectively.It also achieves the highest CO3Dv2 AUC-30 among evaluated non-dense methods for each backbone and matches dense 7Scenes accuracy errors at reported precision.Source: §4.2 · p. 7, Table 1 · p. 7
  • At approximately 94% mean attention sparsity, trained Zebra3D-VGGT reaches 0.805 Re10K AUC@30, within 1.1 percentage points of dense VGGT fine-tuned on the same data and schedule.For VGGT, the complete objective raises Re10K AUC@30 from 0.737 to 0.805; training also reduces π3's aggressive-configuration NRGBD accuracy error from 0.043 to 0.017.Source: §4.2 · p. 7
  • At 1,000 views, Zebra3D-KD (0.7L/0.6S) achieves a 0.447 Chamfer distance versus dense VGGT's 0.466, while the 0.7L/0.8S variant reaches 5.6× speedup with a 0.460 Chamfer distance.The 0.7L/0.8S speedup grows from 1.86× at 100 views to 5.60× at 1,000 views.Source: Table 2 · p. 8, §4.2 · p. 9
  • At 1,024 views, Zebra3D (0.7L/0.8S) reaches 6.2×, 5.5×, and 7.1× end-to-end speedups over dense VGGT, HYWM, and π3, respectively, exceeding A3D and Block Sparse 3D.Across Tanks and Temples, ETH3D, and DTU at 8–64 views, all 0.5L configurations remain within 3% of dense backbones in NC1 and outperform A3D.Source: Figure 5 · p. 8, §4.2 · p. 9

4.3 Ablation Studies

  • Sensitivity-guided placement achieves the highest Re10K AUC@30 on all three backbones at a fixed 0.5L/0.8S budget, while uniform, random, and reverse-sensitivity assignments degrade substantially.Across five calibration datasets, sensitivity rankings have Spearman correlations of 0.843–0.886, and allocations agree on 81.5–88.9% of stages at 0.5L and 93.5–97.2% at 0.7L.Source: §4.3 · p. 10
  • On π3, affine fitting raises Re10K AUC@30 from 0.789 to 0.811 and centering to 0.836; landmark summaries raise AUC@30 from 0.811 to 0.868 and reduce NRGBD accuracy error from 0.066 to 0.014.Landmarks summarize unselected tiles, adding compressed context beyond exact attention over selected tokens.Source: Table 3 · p. 9, §4.3 · p. 10
  • The complete training objective performs best across all three reported metrics, reaching 0.805 AUC@30 and reducing 7Scenes and NRGBD accuracy errors to 0.015 and 0.024.KD-only training underperforms GT-only training on Re10K (0.716 vs. 0.777 AUC@30), and output KD alone does not close the gap.Source: Table 3 · p. 9, §4.3 · p. 10

5 Conclusion

Zebra3D upcycles pretrained 3D reconstruction models into sparse–linear hybrids by matching attention operators to stage sensitivity. The paper reports close-to-dense accuracy at high sparsity, while identifying privacy-sensitive deployment and future hardware and allocation work as boundaries.

5 Conclusion

  • Zebra3D assigns tile-sparse attention to stages sensitive to linearization and linear-centered attention to tolerant stages, with landmark summaries for omitted context.Closed-form affine fitting enables training-free conversion; optional post-conversion training adds task supervision and output and relational feature distillation.Source: §5 · p. 10

AI Use Statement

  • The authors report using generative AI for language editing and preparing and debugging code and tables, and say they reviewed and verified the assisted content.They take responsibility for the paper’s claims, results, and presentation.Source: AI Use Statement · p. 10

Ethics Statement

  • The work uses existing pretrained models and standard research datasets, with no human-subject experiments or newly collected personal data.The authors note that scalable 3D reconstruction can nevertheless enable privacy-sensitive or surveillance applications, and call for respect for consent, dataset licenses, privacy, and applicable law.Source: Ethics Statement · pp. 10–11

Reproducibility Statement

  • The paper and appendix specify conversions, layer assignments, calibration and training procedures, evaluation protocols, and ablations for reproducing the reported results.The authors state that code, configurations, custom kernels, and trained checkpoints will be released upon publication.Source: Reproducibility Statement · p. 11

A.1 Inference in 3D Reconstruction Models

  • At fixed resolution, global-attention computation grows quadratically with view count, and the dense attention matrix supports long-range correspondence and global geometric consistency.The inference pipeline alternates within-view processing and scene-level mixing before task-specific heads decode camera parameters and dense outputs.Source: §A.1 · p. 14

A.2 Extended Related Work

  • Prior approaches reduce token count, compress key/value context, or replace attention, whereas Zebra3D profiles each pretrained stage and selects linear or tile-sparse attention by sensitivity.Tile-sparse stages retain exact softmax interactions for selected regions and represent omitted regions with landmarks; linear-centered stages retain inherited projections and use a closed-form fit.Source: §A.2 · p. 15

A.3 Comparison with Direct Multi-View Reconstruction Models

Zebra3D compares favorably with direct multi-view reconstruction models across reported benchmarks and offers distinct long-sequence quality–speed operating points. The section also documents conversion, baseline, and calibration procedures.

A.3 Comparison with Direct Multi-View Reconstruction Models

  • 0.871 RE10K AUC-30 and 0.900 CO3Dv2 AUC-30: Zebra3D-π3 gives the strongest overall pose results, while Zebra3D 0.5L/0.6S backbones outperform direct multi-view models on all reported 7Scenes and NRGBD point-map metrics and CO3Dv2 AUC thresholds.On RE10K, Zebra3D-HYWM and Zebra3D-Pi3 exceed MUSt3R in AUC-30 and RTA-30, while MUSt3R remains competitive in RRA-30.Source: §A.3 · p. 15, Table 5 · p. 16

B.1 Backbones and conversion

  • The conversion ranks global stages by single-stage linear-conversion error, assigns the requested fraction to linear-centered attention, and replaces every global-attention stage with a hybrid operator.Table 6 lists the resulting zero-based 0.5L and 0.7L assignments; in the xL/yS shorthand, the selector retains key-tile fraction ρ = 1 − y.Source: §B.1 · p. 16, Table 6 · p. 17

A.4 ScanNet-50 Quality–Speed Trade-off

  • On ScanNet-50 at 1,000 frames, Zebra3D-KD 0.7L/0.6S achieves the lowest Chamfer distance (0.447) at 3.54× speedup; 0.7L/0.8S reaches 5.60× speedup with CD 0.460.Source: §A.4 · p. 16, Figure 7 · p. 17

B.2 Baselines

  • Baselines include dense backbones and published or architecture-matched efficient-attention methods, evaluated through the same data loaders, frame selections, alignment, and metric code as Zebra3D.A3D reimplements AVGGT-f4 on VGGT and ports its operator unchanged to HYWM and π3; unsupported backbone–baseline combinations are omitted.Source: §B.2 · pp. 17–18

B.3 Linear initialization

  • Training-free linear-centered stages are calibrated on 13 held-out multi-view scenes with a regularized float64 least-squares fit, without gradient updates to inherited weights.The fit uses teacher softmax outputs and converted linear outputs accumulated across scenes, tokens, heads, and views.Source: §B.3 · p. 18

B.4 Triton implementation of hybrid attention

The implementation uses packed, tiled sparse attention with landmark summaries and fused GPU kernels, alongside a training bridge for refinement. Relational distillation trains converted students against dense teachers using output and feature-based losses.

B.4 Triton implementation of hybrid attention

  • Tile-sparse attention partitions each image into 8×8-patch tiles, selects key blocks per head and query block, and represents unselected blocks with key–value landmarks.Special-token keys remain visible, special-token query rows retain dense attention, and each patch query retains its self key.Source: §B.4 · p. 18
  • Cached layout permutations and fused Q/K/V packing support coalesced access, while linear-centered kernels fuse Q/K normalization, positional rotation, and feature mapping.These implementation choices avoid repeated layout-dependent index construction and intermediate normalized Q/K tensors.Source: §B.4 · p. 18
  • A custom autograd bridge computes sparse-attention gradients for training, while inference calls the forward kernel directly.The sparse call is an explicit compiler graph break; the reported 3,000-step KD runs use the bridge.Source: §B.4 · p. 19

B.5 Relational knowledge distillation

  • Each Zebra3D-KD student starts from a training-free conversion and is refined for 3,000 steps against a frozen FP32 original-checkpoint teacher.Output matching covers modalities shared by teacher and student; the distillation weight decays linearly from 1.0 to 0.6 during refinement.Source: §B.5 · pp. 19–20
  • Feature-L1 taps use backbone-specific intermediate outputs, while cosine and relational matching sample four linear-centered stages per step and compare teacher-selected cross-view token pairs.The relation loss uses 192 anchors per sampled stage, retains the teacher’s 8 most similar candidates from each pool, and adds 24 further candidates.Source: §B.5 · pp. 19–20

B.6 Training data and optimization

The appendix specifies the training data, objectives, optimization, evaluation protocols, metrics, runtime measurements, and sparsity calculations used in the study.

B.6 Training data and optimization

  • KD uses a 20% subset of the HunyuanWorld-Mirror training mixture, while training-free conversion fits projections on 13 calibration scenes without gradient updates.Training sequences contain at most 16 views; evaluation on longer sequences tests generalization beyond that training length.Source: §B.6 · pp. 20–21
  • HYWM and VGGT use shared camera, point-map, and depth objectives, whereas π3 uses its published local-point-map and camera-to-world objective.The dense VGGT control uses the same data, augmentation, training schedule, optimizer, and seed as Zebra3D-KD, but task supervision alone.Source: §B.6 · pp. 20–21

B.7 Evaluation datasets

  • Indoor point-map evaluations use fixed sequence maps, while RealEstate10K and CO3Dv2 pose evaluations use ten-frame clips or sequences.The passage specifies dataset-specific frame selection and preprocessing, including category macro-averaging for CO3Dv2.Source: §B.7 · p. 21

B.8 ScanNet-50 long-sequence protocol

  • ScanNet-50 reconstruction uses 50 scenes and frame-index maps at 1,000, 500, 300, and 100 frames, with Chamfer distance computed on 518×392 inputs.The protocol downsamples and clips the ground-truth mesh, samples 100k points, and aggregates scene medians; Figure 5 uses the runtime harness at 1,024 frames.Source: §B.8 · p. 21

B.9 Metrics and alignment

  • Point-map metrics use similarity alignment followed by ICP, while pose AUC-30 integrates the cumulative distribution of the larger rotation or translation error up to 30°.RealEstate10K reports relative-pose accuracy at 5°, 15°, and 30°; CO3Dv2 reports AUC at 3°, 5°, 15°, and 30°.Source: §B.9 · pp. 21–22

B.10 Runtime measurement

  • Runtime speedup is the separately measured dense median divided by the method median for the same backbone and frame count.Measurements include the model forward pass and enabled geometry heads, but exclude input loading and preprocessing.Source: §B.10 · p. 22

B.11 Mean Attention Sparsity

  • Mean attention sparsity is defined from evaluated query–key pairs relative to dense attention, with special-token terms retained in reported examples and plots.The appendix describes how Zebra3D sparse layers count selected tiles, landmarks, special-token keys, self keys, and dense special-token query rows.Source: §B.11 · p. 22
  • The appendix also specifies attention-pair counts for A3D and RegimeVGGT, including their retained-key and layer-specific token-count conventions.For RegimeVGGT, protected patches can make counts vary across scenes, so dataset results use exact inference counts.Source: §B.11 · pp. 22–23

C Layer Structure and Allocation Analysis

Across backbones, sensitive middle stages combine concentrated attention with greater disruption from linear conversion, motivating stage-specific operator assignment. The calibration-set rankings remain similar across datasets.

C.1 Depth-Wise Structure Across Backbones

  • Across VGGT, HYWM, and π3, global attention dominates long-sequence inference, while a narrow middle band is most disrupted by linear conversion and has concentrated attention.Figure 8 extends the measurements to HYWM and π3; its shaded sensitive bands are stages 11–16 for HYWM and 11–15 for π3.Source: §C.1 · p. 23, Figure 8 · p. 24

C.2 Layer-Role Placement Ablation

  • In the placement study, all rows keep the 0.5L/0.8S budget, linear/tile-sparse stage counts, and closed-form initialization fixed, changing only which stages receive the linear role; no row receives gradient updates.Source: §C.2 · p. 23

C.3 Calibration-Set Stability of Layer Rankings

  • Across five calibration datasets, sensitivity rankings have Spearman’s ρ of 0.7765–0.9365, and assignment agreement is higher at 0.7L than at 0.5L.Agreement is 81.5%–88.9% at 0.5L and 93.5%–97.2% at 0.7L; the assignment is not specific to one calibration collection.Source: §C.3 · p. 23
  • Figure 8 reports oracle-retained attention mass at a 20% block budget and output deviation after one-stage linear-centered replacement, averaged over 11 held-out scenes.The caption places HYWM above π3 and marks each model’s sensitive middle-stage band.Source: Figure 8 · p. 24

D.1 Zero-Shot Linear-Attention Ablation

Zero-shot linear-centered attention performs robustly across the reported pose and point-map comparisons, while VGGT attention concentration changes across depth.

D.1 Zero-Shot Linear-Attention Ablation

  • Linear-centered attention achieves the best RealEstate10K AUC-30 in every backbone–assignment group, with especially large gains over frame attention at 0.7L.Point-map results vary by metric, but fitted linear attention generally avoids frame attention’s severe 0.7L degradation and provides the strongest overall robustness across reconstruction and pose.Source: §D.1 · p. 24
  • Across the reported ablations, linear-centered attention’s pose advantage is accompanied by point-map results that depend on the metric.The comparison covers 7Scenes, Neural RGB-D, and RealEstate10K and CO3Dv2 pose metrics.Source: §D.1 · p. 24
  • VGGT teacher attention becomes concentrated and stripe-structured at stage 11, then exits that pattern after stage 15, matching the 0.7L assignment.The figure spans eight global stages across both linear bands and the tile-sparse band.Source: Figure 9 · p. 25

D.2 Training-Based Linear-Attention Ablation

Training-based mixer comparisons favor linear-centered attention on both pose benchmarks, while tile-sparse attention with landmarks provides the strongest aggregate sparse-attention trade-off. The section also describes the distillation ablation and supplementary result displays.

D.2 Training-Based Linear-Attention Ablation

  • 0.852 RealEstate10K AUC-30 and 0.849 CO3Dv2 AUC-30 are the best pose-benchmark results for linear-centered attention among the compared trained HYWM mixers.No mixer dominates every metric: Sparse Linear Attn performs best on NRGBD, while Gated Delta Net leads several 7Scenes metrics.Source: §D.2 · pp. 25–26

D.3 Sparse-Attention Ablation

  • Tile sparse with landmarks provides the strongest aggregate sparse-attention trade-off, generally outperforming block sparse and surpassing SGA on NRGBD and RealEstate10K.SGA remains competitive on some 7Scenes metrics; the ablation spans all three backbones at 0.5L and 0.7L with 0.75S.Source: §D.3 · p. 26

D.4 Knowledge-Distillation Ablation

  • The VGGT objective ablation compares task-only and distillation-only training with output, feature, and combined relational distillation variants.RelKD combines task supervision with output, feature-L1, token-cosine, and cross-view relational losses.Source: §D.4 · pp. 26–27

E.1 Full Main-Benchmark Results

  • Tables 20 and 21 provide the complete reconstruction and camera-pose metric breakdowns summarized by Table 1.Source: §E.1 · p. 27

E.2 Additional Performance–Sparsity Results

  • Figure 10 reports 300-frame performance–sparsity trade-offs for VGGT and π3 using the same sparsity calculation and plotting protocol as the other backbones.Source: §E.2 · p. 27

E.3 Wide-Baseline Point-Map Reconstruction

The wide-baseline evaluation tests converted models on Tanks and Temples and ETH3D, with DTU as a narrow-baseline control. The 0.5L setting transfers more reliably than 0.7L, especially for π3.

E.3 Wide-Baseline Point-Map Reconstruction

  • The evaluation averages scenes and view counts N ∈ {8, 16, 32, 64} on Tanks and Temples and ETH3D, retaining DTU as a narrow-baseline control.Each converted model is compared with its own dense backbone on the same scenes to isolate conversion cost from backbone quality.Source: §E.3 · p. 27, Table 22 · p. 35
  • At 0.7L/0.8S, π3 loses 0.076 NC1 on Tanks and Temples and 0.054 on ETH3D; its DTU accuracy rises from 1.0550 mm to 1.9583 mm at 0.7L/0.6S.The section links this vulnerability to π3's later sensitive band, which is encroached on sooner as the linear fraction rises.Source: §E.3 · p. 28
  • The authors regard 0.5L as the wide-baseline operating point, while 0.7L is a speed-first setting validated mainly on video-like benchmarks.On Tanks and Temples, HYWM's completeness error at 0.5L/0.6S is +0.53 m versus dense, reduced to +0.04 m after distillation.Source: §E.3 · pp. 28–29
  • Every 0.5L configuration stays within 3% of its dense backbone in NC1 across three benchmarks and backbones; the largest gap is 2.9% (0.6790 versus 0.6992).On ETH3D, every VGGT 0.5L configuration also beats dense VGGT in accuracy and completeness; Zebra3D-KD 0.5L/0.6S records 0.2438 m accuracy and 0.1816 m completeness, versus 0.2898 m and 0.3321 m for dense VGGT.Source: §E.3 · p. 28, Table 22 · p. 35

F.1 Additional Qualitative Reconstructions

The qualitative reconstructions include ScanNet-50 scenes and scenes outside the training data, while a quantitative analysis finds improvements across point-error quantiles without thicker surfaces. The paper also identifies boundaries on where Zebra3D's speedups and allocation strategy apply.

F.1 Additional Qualitative Reconstructions

  • Figure 11 shows four ScanNet-50 scenes colored by distance to ground truth; scenes were selected by recorded per-scene Chamfer distance before rendering.The render protocol refits the same similarity transform for every method, removes no points, and uses the evaluation's bounding-box alignment for reported Chamfer distances.Source: Figure 11 · p. 36, §F.1 · p. 36
  • Across four ScanNet-50 scenes, Zebra3D-KD (0.7L/0.6S) reduces median per-point error from 0.217 to 0.174 m and points beyond 0.30 m from 33.0% to 23.1%.The 90th-percentile error falls from 0.536 to 0.420 m, while estimated local surface thickness is 0.0114 m versus 0.0112 m for dense VGGT; more predicted depth clears the shared confidence threshold.Source: §F.1 · p. 36
  • Recomputed Chamfer distances reproduce recorded values to ±0.0000 for dense VGGT, A3D, and Zebra3D-KD; FastVGGT and RegimeVGGT vary by up to 0.005, without changing Table 2's ordering.The stated explanation is that their token-merging decisions depend on runtime state and are not bit-reproducible across runs.Source: §F.1 · p. 36
  • Figure 12 presents an outdoor KITTI sequence and three Neural RGB-D rooms, none in the training data of any model shown.The comparison is qualitative, with image-colored point clouds and the same lowest-confidence 40% and farthest 3% of pixels removed for each method.Source: §F.1 · p. 36, Figure 12 · p. 37

F.2 Limitations

  • Zebra3D's speedups are modest at 8–16 views, and its fixed per-backbone allocation neither models interactions among converted stages nor adapts to individual inputs.Generalization to substantially different architectures and domains remains to be established; optimized kernels and benchmarks target Triton on AMD MI355X GPUs.Source: §F.2 · p. 37

Related papers

Loading 2610.12727v1…