PaperlayerView exact revision on arXiv

Source-linked AI summary

T3lescope: Arbitrary-Resolution High-Fidelity Generative Surface Reconstruction from Images

Atsuhiro Noguchi, Tianhan Xu, Yiming Liang, Yuta Kikuchi, Masahiro Ishiyama, Shintaro Takagi, Hitoshi Murai, Eiichi Matsumoto

arXiv:2610.03308v1cs.CV

TL;DR

Per-scene optimization lacks a learned 3D prior for sparse or challenging surfaces, while fixed-resolution generators trade spatial coverage against detail. T3lescope reuses one generator in an inference-time coarse-to-fine cascade and outperforms feed-forward and generative baselines while matching or surpassing per-scene optimization across scene types. Its reported results generalize across scenes, view counts, and image resolutions.

  • Problem

    Per-scene optimization lacks a learned 3D prior for sparse views and challenging surfaces, while fixed-resolution generators trade spatial coverage against geometric detail.

    Source: §1 · p. 2

  • Method

    T3lescope reuses one fixed-resolution generator across an inference-time coarse-to-fine cascade, with shared weights and cell scales, layout, and depth chosen at inference time.

    Source: §1 · p. 2

  • Results

    T3lescope outperforms feed-forward and generative baselines and is competitive with per-scene optimization across indoor, outdoor, and city-scale scenes.

    Source: §1 · pp. 2–3, §5 · p. 9

  • Takeaways & Limitations

    The reported results show generalization across diverse scenes, view counts, and image resolutions.

    Source: §1 · p. 2

    High-resolution sampling of large scenes remains slow as cost grows with cells and levels; the method assumes posed images and does not work on distant objects in unbounded scenes.

    Source: §5 · p. 10

Abstract

from arXiv · show

We reconstruct high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization, across scales ranging from single objects to large outdoor scenes. Per-scene optimization methods lack the learned 3D prior needed when observations are sparse or surfaces are glossy or transparent. Existing generative methods leverage such priors to complete geometry in sparsely observed regions, but typically operate at a fixed resolution over a limited spatial extent, trading spatial coverage against detail. Reconstructing a large scene therefore often requires partitioning it into independently processed overlapping local regions, making it difficult to maintain global geometric consistency. To address these issues, we propose T3lescope, which applies a single fixed-resolution generator across scene scales in an inference-time coarse-to-fine cascade. A coarse level establishes the scene layout, and finer levels perturb and denoise geometry inherited from the coarser level within progressively finer spatial cells to recover surface detail. The model is trained on individual cells at multiple scales and shares its weights across all levels, so no hierarchy is fixed during training, and the number of levels, cell scales, and cell locations are determined at inference time. On indoor, outdoor, and city-scale scenes, T3lescope outperforms feed-forward and generative baselines, matches or surpasses per-scene optimization, and recovers fine structures as well as glossy and transparent surfaces. These results show that our method generalizes across diverse scenes, view counts, and image resolutions. Project page: https://pfnet-research.github.io/t3lescope/

1 INTRODUCTION

T3lescope targets scene reconstruction settings where fixed-scale generators trade spatial coverage against detail and overlapping local regions make global consistency difficult. It reuses one generator in an inference-time coarse-to-fine cascade, and performs strongly across indoor, outdoor, and city-scale evaluations.

1 INTRODUCTION

  • Existing generators trade spatial coverage against geometric detail, while overlapping local cells interact indirectly and make global geometry difficult to maintain.Per-scene optimization also lacks a learned 3D prior for sparse views, occlusion, and challenging glossy, reflective, or transparent surfaces.Source: §1 · p. 2
  • T3lescope reuses one fixed-resolution generator in an inference-time cascade, with finer levels perturbing and denoising inherited geometry in smaller cells.Cell size, layout, and cascade depth determine coverage and output resolution at inference time.Source: §1 · p. 2
  • Its cell-independent multi-scale conditioning uses reusable image features to support cells at arbitrary positions and scales.Each voxel is conditioned using its projected location and size, providing image detail suited to each cascade level.Source: §1 · p. 2
  • T3lescope outperforms feed-forward and generative baselines and is competitive with per-scene optimization on indoor, outdoor, and city-scale scenes.Evaluations include ScanNet++, Tanks and Temples, and a city-block dataset, under sparse- and dense-view settings.Source: §1 · pp. 2–3

2 RELATED WORK

Prior reconstruction approaches trade optimization cost, learned scene priors, and cross-view consistency in different ways. T3lescope instead propagates global structure through a shared 3D generator whose hierarchy is configured at inference time.

2 RELATED WORK

  • Per-scene optimization can recover accurate geometry from dense views but requires scene-specific fitting and degrades with sparse observations or glossy and transparent surfaces.Feed-forward methods avoid per-scene optimization but fuse image-aligned predictions, making meshes sensitive to cross-view inconsistencies.Source: §2 · p. 3
  • GenRecon extends a fixed-resolution generator with overlapping cells, but distant regions interact only indirectly through neighboring overlaps.T3lescope instead propagates global structure from the coarse reconstruction to each finer cell.Source: §2 · p. 3
  • Earlier coarse-to-fine 3D generators fix their hierarchy during training, whereas T3lescope reuses one model across scales and configures the hierarchy at inference time.At finer levels, it reduces cell size while keeping generator resolution fixed.Source: §2 · p. 3

3 METHOD

T3lescope generates a scene with a shared, fixed-resolution model applied to overlapping cells, then refines inherited geometry through an inference-time hierarchy. It trains across cell scales and conditions cells using reusable, scale-matched image features.

3 METHOD

  • The pipeline applies one generator with corresponding image features to overlapping cells, trains a single model across scales, and configures the hierarchy at inference time.The coarsest level captures the scene; finer levels refine it.Source: §3 · p. 3

3.1 SINGLE-LEVEL SCENE GENERATION

  • The single-level generator creates sparse surface voxels before generating a mesh, and overlapping cells share a scene-wide latent updated through MultiDiffusion.Cells use canonical coordinates and can be placed at arbitrary world locations and scales because the model receives neither absolute position nor metric cell size.Source: §3.1 · pp. 3–4

3.2 INFERENCE-TIME COARSE-TO-FINE CASCADE

  • Each cascade level halves cell size; finer cells derive a noisy sparse-structure initialization from the parent mesh and refine it with the SS flow.The noising level trades preservation of parent geometry against incorporation of finer image detail; the authors set it to 0.8 across benchmarks without per-scene, dataset, or view-count tuning.Source: §3.2 · pp. 4–5

3.3 MULTI-SCALE TRAINING OF A SHARED MODEL

  • One pair of flow models is trained on individual cells at multiple scales without paired coarse-to-fine examples, leaving hierarchy depth, cell layout, and physical scales configurable at inference time.Cells map to the same canonical coordinates and grid; their voxel spacing differs, but scale is not given to the model.Source: §3.3 · p. 5

3.4 CELL-INDEPENDENT MULTI-SCALE IMAGE CONDITIONING

  • Voxel-queried attention lets current voxel features attend to projected image tokens, while a tile bank supplies scale-matched features that can be reused across cells and flow steps.Images are encoded as a pyramid of fixed-size tiles; each cell selects tiles matching its voxel spacing. Figure 2 summarizes the cell refinement, tile selection, and shared generators.Source: Figure 2 · p. 4, §3.4 · pp. 5–6

4 EXPERIMENTS

Experiments compare T3lescope with optimization, feed-forward, and generative baselines across indoor, outdoor, and city-block scenes. Results favor T3lescope against generative baselines, with parent initialization and image conditioning both examined in ablations.

4.1 EXPERIMENTAL SETUP

  • The general reconstruction model uses synthetic training data, while the city-block model is fine-tuned on Japanese city photographs paired with BIMs.Both models are initialized from pretrained TRELLIS.2 and fine-tuned for multi-view conditioning on TexVerse.Source: §4.1 · p. 6
  • Evaluation covers ScanNet++, Tanks and Temples, and three held-out city blocks, with per-scene optimization, feed-forward, and generative baselines.The reported metrics include Chamfer distance, F-score, normal consistency, and held-out-view 2D measures on ScanNet++.Source: §4.1 · p. 7

4.2 COMPARISON WITH PRIOR RECONSTRUCTION METHODS

  • On ScanNet++, the finest level beats GenRecon on all seven metrics at 32 and 256 views, and beats optimization baselines on all three 3D metrics.DA3 has higher F@2 at both view counts and slightly lower Chamfer distance plus higher completeness at 256 views; T3lescope's accuracy changes little between view counts.Source: §4.2 · pp. 7–8, Table 1 · p. 7
  • On Tanks and Temples, T3lescope achieves the lowest or a competitive Chamfer distance in all settings and the highest F-score on large scenes with 16 views.Its surfaces can be slightly offset from ground truth, which tight thresholds of 3 to 25 mm penalize.Source: §4.2 · p. 8, Table 2 · p. 8
  • On city blocks, baselines have Chamfer distances of about 5 m or more, while T3lescope reaches 0.65 m and outperforms them on all three metrics.Each finer cascade level improves all three metrics.Source: §4.2 · p. 8, Table 2 · p. 8

4.3 ABLATIONS AND ANALYSIS

  • Parent initialization improves every metric on large ScanNet++ scenes with 32 views, and lowers Chamfer distance by 80%.With parent initialization, neighboring cells preserve the coarse building roof; without it, the roof breaks into fragments.Source: §4.3 · p. 9, Figure 5 · p. 9
  • On ScanNet++, adding the ConvNeXt branch to voxel-queried attention substantially improves every metric over the baseline conditioner.In the human capture, voxel-queried attention recovers hair strands, while ConvNeXt sharpens them and adds shirt buttons.Source: §4.3 · p. 9, Figure 6 · p. 9
  • On large ScanNet++ scenes, inference takes 17 minutes at L1 and 89 minutes at L2 on one H200, comparable to optimization but slower than feed-forward methods.The times are reported per scene.Source: §4.3 · p. 9

5 CONCLUSION AND LIMITATIONS

T3lescope uses a shared fixed-resolution generator in an inference-time cascade, with coarse levels establishing layout and finer levels recovering detail. The authors report competitive or better results against reconstruction baselines, while noting limits for unbounded scenes, runtime, and image alignment.

5 CONCLUSION AND LIMITATIONS

  • T3lescope applies one fixed-resolution 3D generator across scene scales through an inference-time coarse-to-fine cascade.The coarse level establishes globally consistent layout, finer levels recover surface detail, and cell-independent multi-scale conditioning supplies images at each cell's needed resolution.Source: §5 · p. 9
  • On ScanNet++, Tanks and Temples, and city blocks, T3lescope outperforms the generative baseline and is competitive with or better than feed-forward and per-scene optimization methods.Source: §5 · p. 9
  • T3lescope assumes posed images, does not reconstruct distant objects in unbounded scenes, and becomes slow when large scenes are sampled at high resolution.Its cost grows with the number of cells and cascade levels; surface offsets can also lower F-score at tight thresholds when views are dense.Source: §5 · p. 10

A ADDITIONAL RESULTS

The appendix expands benchmark results and qualitative comparisons, describes evaluation data and metrics, and documents a failure case for distant objects in unbounded scenes.

A.1 SCANNET++

  • DA3 has the lowest MAE, AbsRel, and Chamfer distance on small ScanNet++ scenes, while T3lescope’s finest level has the lowest normal error in all four settings.Tables 3 and 4 cover small, large, and all 50 validation scenes under 32- and 256-view settings.Source: Table 3 · p. 14, Table 4 · p. 14, §A.1 · p. 14

A.2 TANKS AND TEMPLES

  • With 4K inputs, the finest level achieves higher mean F@τ in every view setting, including 0.3344 vs. 0.3009 with 256 views, and lower CD/τ at 256 views.The gains are concentrated at finer levels, whose cells read full-resolution tiles.Source: Table 5 · p. 20, §A.2 · p. 20

A.3 CITY BLOCKS

  • Each finer cascade level improves all three reported metrics on every city-block evaluation block.Table 6 reports per-block results with 256 input views.Source: Table 6 · p. 26, §A.3 · p. 26

A.4 COMPARISON WITH GAUSSIANWRAPPING ON REAL-WORLD CAPTURES

  • On garden and an in-house five-person capture, GaussianWrapping and T3lescope use the same views and render camera, but GaussianWrapping receives 2K images.Figure 17 shows the real-world mesh comparison with matching shading.Source: §A.4 · p. 26, Figure 17 · p. 27

A.5 FAILURE CASE IN UNBOUNDED SCENES

  • Even with four levels and cells placed up to eight times farther from the scene center than the cameras, distant ground and trees are not reconstructed well; infinite-distance objects are unsupported by construction.The authors suggest limited training-data diversity as a possible cause of the distant-object failure.Source: §A.5 · p. 28, Figure 18 · p. 28

B.1 DATASETS

  • Evaluation covers 50 ScanNet++ validation scenes, six Tanks and Temples scenes, and three held-out city blocks from a 41-block dataset.The city-block dataset contains 248 buildings and 139,893 posed images.Source: §B.1 · p. 29

B.2 METRICS

  • The reported metrics include mesh-based Chamfer distance, F-score and normal consistency, plus rendered depth MAE and AbsRel and normal angular error.F@τ is the F-score at threshold τ; depth and normal errors are computed from rendered maps.Source: §B.2 · p. 29

B.3 EVALUATION PROTOCOLS

The appendix specifies dataset-specific evaluation protocols and baseline implementations, then details the inference pipeline and its two-stage 3D flow generator.

B.3 EVALUATION PROTOCOLS

  • ScanNet++ ground truth is rigidly aligned to COLMAP sparse points with ICP, while 2D errors are pooled over valid pixels from official test-split views.The 2D evaluation reimplements GenRecon’s protocol because its evaluation code is unreleased.Source: §B.3 · p. 30
  • The virtual-scan protocol scores camera-visible ground-truth surfaces; Tanks and Temples uses trajectory cameras, official crops, ICP alignment, and scene-specific F-score thresholds.For city blocks, evaluation compares complete meshes, cropping reconstructions to the ground-truth footprint.Source: §B.3 · p. 30

B.4 BASELINES

  • Baselines receive the same views, poses, and intrinsics, but image resolution and mesh extraction vary; GenRecon generates overlapping chunks and aggregates their latents.Depth-map methods are fused with TSDF, while MILo, GaussianWrapping, and GenRecon output meshes directly.Source: §B.4 · pp. 30–31

C ADDITIONAL METHOD DETAILS

  • At each cascade level, cells cover the scene or parent surface, select conditioning views and tiles, and run scene-wide occupancy and mesh stages; parent subdivision initializes finer levels.The procedure decodes the finest-level shape latent into a mesh, with other levels optionally decoded.Source: C ADDITIONAL METHOD DETAILS · p. 31

C.1 TWO-STAGE 3D FLOW BACKBONE

  • The generator first predicts 64^3 occupancy, then generates an occupancy-conditioned shape latent decoded into a 512^3 mesh; its pretrained VAEs remain frozen during flow-model training.The occupancy and mesh stages use separate latent flow-matching models.Source: §C.1 · pp. 31–32

C.2 MULTI-SCALE TRAINING WITH SHARED WEIGHTS

The method shares one generator across cell scales, blends overlapping predictions in a scene-wide latent grid, and uses parent geometry to guide finer cells and their placements.

C.2 MULTI-SCALE TRAINING WITH SHARED WEIGHTS

  • Training maps cells of different physical sizes into identical canonical coordinates and grids, sharing one weight set per stage while training on individual cells rather than parent–child pairs.Cell size is sampled explicitly from two training ladders to avoid bias toward smaller cells.Source: §C.2 · p. 32

C.3 SCENE-WIDE MULTIDIFFUSION

  • Cells overlap at a half-cell stride, and their latent predictions are blended voxelwise with Hann-window weights that reduce boundary influence.The same weighted averaging applies to dense occupancy and sparse shape latent grids.Source: §C.3 · pp. 32–33

C.4 PARENT-TO-CHILD GEOMETRY TRANSFER

  • Child sparse structure comes from subdividing the parent shape latent; parent occupancy also restricts child placement, with a one-voxel dilation allowing narrowly missed regions to be recovered.Starting the finer occupancy flow from a noised parent latent requires fewer Euler steps.Source: §C.4 · p. 33

C.5 CELL SIZES AND EXTENTS

  • Cell sizes are specified in meters for ScanNet++ and city blocks, but in normalized units for Tanks and Temples and the real-world captures; coarse extents use dataset-specific bounds.For garden, nested cascade levels cover squares with halfwidths of 16, 8, and 4 units around the camera centroid.Source: §C.5 · p. 33

D.1 VOXEL-QUERIED IMAGE ATTENTION

Voxel-queried attention conditions each voxel on image features while addressing the resolution and view-aggregation limits of pixel-aligned conditioning. A cell-independent tile bank and DA3-based layout and view selection support this conditioning across cells and scales.

D.1 VOXEL-QUERIED IMAGE ATTENTION

  • Offset-aware voxel queries preserve sub-token image position without upsampling, while shallow ConvNeXt features supply finer detail at about 1% of upsampled DINO feature memory.Both feature types fit in memory across views and are reused across cells and flow steps; view weights depend on the voxel feature.Source: §D.1 · pp. 34–35
  • Attention encodes viewing-ray direction, pixel offsets, and apparent voxel size so inputs depend on how a cell appears in images, not its absolute position or cascade level.These relative quantities are intended to let the same attention mechanism operate across cell layouts and scales.Source: §D.1 · p. 35
  • The DINO branch jointly attends over nearby tokens and valid views, while the ConvNeXt branch attends over views only.The DINO window selects four tokens around each voxel projection; attention weights vary with features rather than relying on fixed bilinear weights.Source: §D.1 · pp. 34–36

D.2 MULTI-SCALE TILE FEATURE BANK

  • A cell-independent tile pyramid lets cells retrieve appropriately scaled image crops without separately cropping and re-encoding shared image regions.The feature bank stores tile features for reuse across overlapping cells and flow steps.Source: §D.2 · pp. 36–37

D.3 DA3-BASED LAYOUT AND VIEW SELECTION

  • DA3 depth guides coarsest cell placement and derives candidate cells at finer levels from parent occupancy, dilated to avoid losing regions after small coarse-level misses.DA3 is used for these layout decisions, not fed to the flow models or fused into or used to optimize the output mesh.Source: §D.3 · p. 37
  • Tile selection scores visible views by surface-facing angle and pyramid-scale match, then retains each cell’s highest-scoring tiles, up to max(64, N).The score combines the viewing-ray angle with deviation from the preferred image-pyramid level.Source: §D.3 · p. 38

E TRAINING DETAILS

Training adapts pretrained flows to multi-view inputs, then fine-tunes general-scene and city-scale models on different scene data. Cell augmentations span scales and orientations; mesh decoding and cascade merging produce unified scene geometry.

E TRAINING DETAILS

  • Both models start from TRELLIS.2 weights; the mesh flow is pruned from about 1.3B to 655M parameters, then fine-tuned, while the pretrained VAEs remain frozen.Only the two flow networks and their conditioning layers are updated during training.Source: §E.1 · p. 38, E TRAINING DETAILS · p. 38

E.2 SHARED MULTI-VIEW PRETRAINING

  • Shared TexVerse multi-view pretraining adapts the generators to posed views through a curriculum over mesh resolution and conditioning-view count.The mesh flow progresses from 4 views at resolution 256^3 to 16 views at 512^3; the SS flow progresses from 4 to 16 views.Source: §E.2 · p. 38

E.3 GENERAL-SCENE FINE-TUNING

  • General-scene fine-tuning adds procedural multi-room houses and mesh-based grass scenes to single-room datasets, addressing missing multi-room structure and over-smoothed grass.The procedural data includes 200 connected-room houses and 100 outdoor scenes with 32 m extent.Source: §E.3 · pp. 38–39

E.4 CITY-SCALE FINE-TUNING

  • The city-scale model uses the same representations, conditioning, multi-scale cell training, and inference procedure, but fine-tunes on city-block data paired with BIM-derived mesh targets.Source: §E.4 · p. 39

E.5 SCALE AND ROTATION AUGMENTATION OF CELLS

  • Training varies cell scale and orientation, balancing sampling across room and house cell sizes and applying independent rotations and reflections to cells.The discrete transforms are applied to geometry and camera poses, and reflections also flip images.Source: §E.5 · pp. 39–40

E.6 OPTIMIZATION SETTINGS

  • Both flows use AdamW and BF16 mixed precision; scene-level fine-tuning runs for 25,000 steps per flow on 8 H200 GPUs.Source: §E.6 · p. 40

F.1 MARGIN-BASED CHUNK DECODING AND STITCHING

  • Cells are decoded in blocks with neighboring latent margins, then boundary vertices are merged to form one mesh without duplicate faces or open seams.The margin preserves surrounding context near block boundaries while each block outputs only interior faces.Source: §F.1 · p. 40

F.2 COMBINING CASCADE LEVELS

  • For nonuniform reconstructions, successive cascade levels cover smaller regions and their meshes are merged by replacing coarse faces inside each finer-level boundary.This ring composition is used for qualitative reconstructions; benchmark scenes use identical extents across levels and are evaluated at the finest level.Source: §F.2 · pp. 40–41

G.1 PARENT INITIALIZATION

Parent initialization improves consistency, especially in wide scenes with sparse views, though denoising from pure noise can yield higher F-scores in denser settings. Voxel-queried attention and ConvNeXt improve image conditioning, with the latter sharpening recovered details.

G.1 PARENT INITIALIZATION

  • On large ScanNet++ scenes with 32 views, ts = 0.8 raises F@2 from 0.6306 to 0.7155 and reduces Chamfer distance by 5.0× versus ts = 1.Parent initialization is most effective in wide scenes with sparse views.Source: §G.1 · p. 41, Table 7 · p. 41
  • In denser ScanNet++ settings, ts = 1 reaches higher F@2, but ts = 0.8 lowers Chamfer distance by 2.9× on small scenes with 32 views and 1.6× on large scenes with 256 views; NC is higher in every setting.Source: §G.1 · p. 41, Table 7 · p. 41
  • On Tanks and Temples with 16 views at 4K, ts = 0.8 yields lower F@τ than ts = 1 (0.2756 vs. 0.2830), but lower CD/τ (9.69 vs. 12.01).At ts = 1, the reconstructions show floaters and ground holes.Source: §G.1 · p. 41, Table 8 · p. 41, Figure 20 · p. 42

G.2 IMAGE CONDITIONING

  • On Tanks and Temples with 32 views, voxel-queried attention slightly improves mean F@τ over the baseline at every cascade level, and adding ConvNeXt yields the highest mean F@τ across levels.Qualitatively, the baseline smooths details that voxel-queried attention recovers, while ConvNeXt sharpens them further.Source: §G.2 · p. 42, Table 9 · p. 42, Figure 21 · p. 43

G.3 TILE PYRAMID LEVEL SELECTION

Adaptive tile-pyramid selection performs well across ScanNet++ and Tanks and Temples, while fixed coarse or full-resolution choices have dataset-dependent outcomes. A smaller attention window performs comparably and speeds mesh-stage sampling.

G.3 TILE PYRAMID LEVEL SELECTION

  • On ScanNet++, adaptive selection achieves the lowest MAE across cascade levels and scene groups; Coarse is competitive at levels 1 and 2, while Fine underperforms in nearly every column.The table reports means for the 25 smallest, 25 largest, and all 50 scenes.Source: §G.3 · p. 43, Table 10 · p. 43
  • On Tanks and Temples, Coarse reaches mean F@τ 0.2305 at level 2 versus 0.2937 for adaptive selection, which has the best mean F@τ and CD/τ across levels.Coarse downsamples the roughly 2180-pixel-high 4K inputs by a factor of four.Source: §G.3 · p. 43, Table 11 · p. 44

G.4 ATTENTION WINDOW SIZE

  • Window size w = 2 performs comparably to w = 6 across cascade levels and view counts, while using four rather than 36 tokens per view and speeding mesh-stage sampling by roughly 4×.The comparison is on ScanNet++ with 32 and 256 input views.Source: §G.4 · p. 44, Table 12 · p. 44

G.5 SYNTHETIC TRAINING DATA

Removing synthetic multi-room-house and grass training scenes weakens the finest-level reconstructions of multi-room walls and grassy ground.

G.5 SYNTHETIC TRAINING DATA

  • Without synthetic multi-room houses, the model misses the left and far walls of the shown ScanNet++ scene; the full model reconstructs them.The comparison uses 32 input views and finest-level reconstructions.Source: §G.5 · p. 44, Figure 22 · p. 45
  • Without synthetic grass scenes, the model produces flat, featureless ground on caterpillar, while the full model captures individual grass tufts.The figure caption describes the grass around the loader as becoming a flat surface without synthetic data.Source: §G.5 · p. 44, Figure 22 · p. 45
  • The comparison uses w = 6 rather than the final w = 2; the authors report that the two window sizes perform on par.Source: §G.5 · p. 44

H INFERENCE TIME

Inference time rises across cascade levels, but finer levels allocate cells only where parent levels detect surfaces. On five large ScanNet++ scenes, reconstruction through level 2 averages 89 minutes and peaks at 79 GiB of GPU memory.

H INFERENCE TIME

  • 89 minutes is the average time through level 2, versus 17 minutes through level 1, across five large ScanNet++ scenes.Measurements use 32 input views on one NVIDIA H200 GPU.Source: H INFERENCE TIME · p. 45, Table 13 · p. 45
  • Cell counts rise from 31 to 470 to 2,490; runtime increases 8.2× from level 0 to 1 and 4.7× from level 1 to 2.Each level halves cell size, and finer levels allocate cells only where the parent detects surfaces.Source: H INFERENCE TIME · p. 45, Table 13 · p. 45
  • Flow sampling in the SS and mesh stages accounts for most runtime at every level, while view selection, tile preparation, and parent transfer take under 3 minutes combined.Source: H INFERENCE TIME · p. 45, Table 13 · p. 45
  • Peak GPU memory usage is 79 GiB at the finest cascade level.The measurements use one NVIDIA H200 GPU.Source: H INFERENCE TIME · p. 45, Table 13 · p. 45

Related papers

Loading 2610.03308v1…