Aerial Path Online Planning for Urban Scene Updation

Hui Huang; Jianwei Hu; Ke Xie; Mingfeng Tang; Ningna Wang; Xiaohu Guo; Ziyuan Xie

REVIEW 4 major objections 6 minor 5 references

Aerial Path Online Planning for Urban Scene Updation

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper presents the first aerial path planner built for updating—not re-scanning—urban 3D scenes, and reports that it detects changed regions and reconstructs them with accuracy close to full-scene re-exploration at a fraction of the…

desk verdict A genuinely new systems contribution — change-aware UAV path planning for urban scene updates — with plausible large efficiency gains, but the load-bearing change-detection rule is never validated on its own, so treat the headline numbers as conditional. read the letter →

arxiv 2505.01486 v3 pith:ZZ3HSHNS submitted 2025-05-02 cs.RO cs.GR

Mingfeng Tang , Ningna Wang , Ziyuan Xie , Jianwei Hu , Ke Xie , Xiaohu Guo , Hui Huang This is my paper · ORCID

classification cs.ROcs.GR

keywords aerialpathplanningurbansceneupdatechangedetectionnext-best-view3DreconstructionUAVexplorationprobabilityprior

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Urban 3D scenes are often re-scanned in full at regular intervals, even though most of a city has not changed. This paper argues that periodic scene updates can instead be treated as a targeted search problem: fly a UAV along a path that prioritizes areas likely to have changed, detect changes in real time by comparing captured images against renderings of the prior model, and then re-route the drone to fully cover each changed region before resuming the original path. If the approach works as claimed, a returning drone can refresh a large urban model in hours rather than days, spending flight time and computation almost entirely on the parts of the scene that actually changed.

What carries the argument

The load-bearing component is the changeability heuristic f(s_i, v_j), which scores each surface sample by its prior change probability multiplied by visibility, with a penalty term that discounts samples already seen by visited views. This single score drives both the prior-path view selection, where redundancy is defined as low expected changeability, and the real-time next-best-view selection, where confirmed change samples receive q(s_i)=1. A second essential component is the detection pipeline that renders T1 images at each view and uses GIM-enhanced DKM feature matching to label unmatched image regions as changed, and DUSt3R to convert those masked regions into a 3D point cloud whose convex hull becomes the target for real-time exploration.

What would settle it

Run the planner on a scene where only the lighting changes between T1 and T2 (e.g., the same buildings, but different sun angle and shadows); if the mismatch detection flags large unchanged areas as changes, the change-detection stage cannot distinguish real structural change from appearance change.

Watch

Extended reading notes

Core claim

The paper's central claim is that change-driven path planning can replace full-scene re-exploration for periodic urban scene updates. Given the T1 reconstructed model plus per-semantic-class change probabilities, the planner first computes a prior path over a reduced set of viewpoints that maximizes expected changeability while guaranteeing coverage. When a captured T2 image and the rendered T1 image fail to match in some region, the system triggers a real-time next-best-view path that expands the detected change area using DUSt3R point clouds and convex hulls. The evaluation reports an IoU of 0.840 and 0.253 for two changed buildings against 0.801 and 0.059 for the stronger grid baseline, with 19 viewpoints and 994 m of flight against 80 viewpoints and 1,121 m, and a 52% trajectory-length reduction and 71% viewpoint reduction compared with re-running a standard full-scene planner on the same five-change scene. Error and completeness of the resulting reconstructions are comparable to a coarse-proxy-based full reconstruction.

Load-bearing premise

The whole trigger mechanism assumes that image regions without feature matches between a rendered old view and a captured new view really are 3D changes, while regions with matches are unchanged.

Editorial extensions

If this is right

If the efficiency numbers hold, periodic urban model refresh—monitoring, digital-twin updates, disaster damage assessment—can be scheduled as short focused flights rather than full re-surveys, with the prior model and historical change statistics doing most of the work.
The framing of a scene-update task as a max-coverage, min-redundancy optimization over a changeability score extends to any setting with a prior model and a class-conditioned change prior, including indoor robotics and infrastructure inspection.
The convex hulls of detected change areas serve directly as input to existing reconstruction planners, so the method integrates with current pipelines rather than requiring a new reconstruction system.
Targeted updates keep the computational load low enough that per-view planning decisions (about 0.06–0.09 s in small scenes, 1–2 s in the large scenes) can run onboard, opening the door to fully autonomous re-survey missions.
Because the real-time path is triggered only by confirmed changes, the method scales with the amount of scene change rather than with the size of the scene, which is the key cost inversion for large urban datasets.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The detection mechanism implies a false-negative asymmetry: changed surfaces that still yield feature matches (e.g., repainted façades or replaced furniture visible from the same angle) would be missed entirely, and nothing in the planner revisits unmatched regions. A testable extension would be to add an explicit confirmation step that revisits low-confidence regions at closer range.
The per-class prior probabilities are the only source of guidance before a change is detected. If those priors are wrong for a particular city (e.g., a district with rapid construction labeled mostly as low-change 'high building'), the prior path will spend its time in the wrong places. The planner's robustness could be measured by degrading the priors and measuring detection recall.
The nearest-eight-image window for DUSt3R reconstruction is the documented cause of fragmented change areas. A natural extension is to widen the window adaptively when the IoU overlap between successive point clouds stays low, trading a little online speed for fewer split convex hulls.
The real-time path takes only the top K=10 candidates and chooses the one closest to the current view, which means the path can be trapped in a local exploration pattern; a longer-horizon lookahead over the change-area boundary would likely reduce over-segmentation further.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is the first path-planning paper I know of that treats urban scene updates as their own problem instead of re-running full-scene reconstruction, and the framing is substantive: the two-stage prior/real-time architecture, with an externally sourced changeability prior driving the initial path and an on-board no-match rule triggering focused exploration, is a sensible division of labor that the cited literature does not contain. The authors are also honest — the supplementary shows a real change area split into two convex hulls, a real-world case with a wrong hull, and a limitations section that names the fixed-altitude problem and DUSt3R's speed.

The headline numbers are plausible and large: 52% shorter path and 71% fewer views than Zhou et al. on a five-change-area scene, with reconstruction error and completeness roughly matched. The reconstruction control is the right comparison, because it shows the convex-hull targets do not degrade the downstream pipeline. One disclosure worth noting: the reported per-step planning time excludes the 10–16 s of GIM/DUSt3R processing per view, which the authors report in the supplementary table; that is honest, but it means 'real-time' describes the planner, not the full loop.

The soft spots, in order. First and most important: the Sec. 5 detection rule ('no GIM match between rendered T1 and captured T2' = change) is load-bearing and never validated on its own. No precision/recall, no failure analysis of the mask; the paper's own failure cases are attributed to DUSt3R's eight-image window, not the mask, so the one component everything depends on is the one least tested. Sun-angle and shadow artifacts would send the drone exploring static areas and erode the 52%; missed matches would silently break the 'comparable to full re-exploration' guarantee. Second, the quantitative comparisons rest on one manually edited scene each, and the IoU for the removed building in Table 1 is 0.253 — better than the grid baseline, but not obviously good. Third, all the hyperparameters (ω, γ, φ, K, the two sampling radii) are hand-set with no sensitivity analysis. Fourth, the β in Eq. (6) of Sec. 4 is undefined — presumably a typo for ω — and the supplementary carries an unfinished placeholder. Minor, but in a camera-ready paper these should not survive.

My read: the method is coherent, the problem is well chosen, and the gap is real, so this deserves serious refereeing. But the efficiency claim is conditional until the mask gets a standalone evaluation and the single-scene comparisons get company. If I were refereeing, I would ask for mask precision/recall and at least one more natural T1/T2 pair before accepting the headline numbers. Worth a slot in the reading group for the methodology discussion alone.

Referee Report

4 major / 6 minor

Summary. This paper proposes a two-stage aerial path planning framework for updating previously reconstructed urban 3D scenes at a revisit time T2. The first stage uses the T1 model, semantic labels, and WUSU prior change statistics to assign a changeability score to surface samples, then greedily selects a sparse prior path that balances coverage against redundancy. During flight, each captured T2 image is compared with a rendered T1 image using GIM; image regions with no feature matches are treated as change masks. DUSt3R reconstructs masked point clouds, convex hulls define change targets, and an online next-best-view planner explores each target before the drone resumes the prior path. Experiments compare against region-division exploration and the full-scene planner of Zhou et al., reporting a 52% shorter trajectory and 71% fewer views in one test scene, comparable reconstruction error and completeness, and additional large-scene and real-world results.

Significance. The paper addresses an under-explored problem---periodic aerial updating of an existing urban reconstruction---and the two-path architecture with a statistical changeability prior is a sensible, potentially reusable design. The authors also make good use of public tools (GIM, DUSt3R, UrbanBIS, WUSU) and include real-world and large-scene tests with documented failure cases. If the no-match change-detection rule is reliable, the reported efficiency gains are substantial. However, the significance is currently conditional: the headline efficiency numbers come from one manually modified scene, and the detection rule itself is not evaluated in isolation, even though it is the sole trigger for real-time replanning and the sole input mask to the reconstruction stage.

major comments (4)

[Sec. 5] The rule that image regions with no GIM feature matches between a rendered T1 image and a captured T2 image are change areas is load-bearing: it triggers the switch to real-time planning and defines the masks passed to DUSt3R. The paper reports no precision/recall evaluation of this rule. Table 1 gives end-to-end convex-hull IoU for only two buildings in one scene (IoU_B1 = 0.840, IoU_B2 = 0.253), and Table 5 reports only counts of detected versus ground-truth change areas; 18 detected versus 17 ground truth in Yuehai is consistent with one false positive plus one miss, one split region, or any of several other outcomes. Since lighting changes, shadows, T1 mesh artifacts, and textureless surfaces can all cause no-match false positives, and since any false positive directly inflates flight time, the authors should either add a quantitative detection-stage evaluation (pixel- or region-level precision/recall on scenes with known changes, including an ablation of the match threshold) or clearly weaken the central efficiency claim.
[Sec. 7.2] The paper's headline efficiency claim---52% shorter path and 71% fewer views---is measured on a single scene with five manually created change areas (Section 7.2, Table 3). The two large UrbanBIS scenes in Table 5, where the scalability argument matters most, are only reported as number of views and path length without any comparison to full-scene re-exploration. To support the abstract's claim of generality, the authors should report the same #Views and Path Len comparison, with a full re-exploration baseline such as Zhou et al., for the large scenes, or explicitly state that the quantitative efficiency comparison is limited to one scene and treat the large-scene results as feasibility demonstrations only.
[Sec. 7.3] The system outputs convex hulls of detected changes, and these hulls are the input to the downstream reconstruction. The supplementary materials acknowledge that a single change area can be fragmented into two convex hulls (Sec. F) and that Case 5 in the real-world scene produces an incorrect hull (Sec. E), attributing both to DUSt3R's reliance on only the nearest eight images. These are not peripheral limitations: they directly affect the quality of the update. The paper should quantify how often such fragmentation occurs across all test changes and how it affects the IoU and reconstruction-quality metrics, rather than presenting only the two best-behaved scenes in the main text. Without this, the phrase 'maintaining high-quality updates comparable to full-scene re-exploration' is not fully supported.
[Eq. (5)] Equation (5) is not a well-posed objective as printed: the same set V_P is written as both the argmin of the redundancy term and the argmax of the coverage term, so the set cannot simultaneously minimize and maximize both expressions. The supplementary's greedy removal procedure (Sec. B.1) is a workable algorithm, but the main text should either define a single scalar objective with explicit weighting or lexicographic ordering, or state the greedy algorithm directly as the definition of the prior path. As written, the formal statement would need to be repaired for the method to be reproducible from the main text alone.

minor comments (6)

[Supplementary Sec. A] The supplementary text contains the incomplete phrase '[geographic area to be completed]' in the description of the WUSU dataset; this placeholder must be filled in before publication.
[Supplementary Table 4] Table 4 maps WUSU labels to UrbanBIS labels but leaves several rows blank (Arable land, Grassland, River, Excavation, Bare surface, Unclassified). The authors should state how these labels are treated in the prior probability computation for Eq. (1), and how zero-change-count labels are handled.
[Supplementary Table 5] The counting of detected versus ground-truth change areas in Table 5 needs a precise matching criterion: the reader cannot tell whether 18 vs. 17 means over-segmentation, a false positive plus a miss, or a different configuration. A short description of the matching rule and a false-positive/false-negative breakdown would make the table interpretable.
[Sec. 7.3] The main text refers to 'Table 6' and 'Fig. 12' for the full reconstruction comparison, but these appear only in the separate supplementary PDF; they should be labeled as supplementary tables and figures to avoid confusion.
[Table 3] In Table 3, 'Ours-Rcon' is used for the reconstruction path, while the text uses 'Ours-Recon' and 'Ours-All'; please unify the naming.
[Sec. 7.1] Supplementary Table 7 reports average GIM and DUSt3R processing times of roughly 10--16 seconds per view, while the main text reports 'real-time path planning time' excluding DUSt3R. Since the real-time path is triggered by changes that are only available after DUSt3R runs, the authors should clarify how the online planning loop remains real-time despite this per-view latency, or explicitly state that the reported real-time figure excludes the detection and reconstruction components.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the planner's inputs are external priors and independent matchers/reconstructors, and the reported gains are not defined by the method's own parameters.

full rationale

The derivation chain is self-contained. The changeability prior is taken from the external WUSU dataset (Table 4) and mapped to UrbanBIS labels; it is not fitted to the test scenes. The Sec. 5 detection rule (no feature matches between rendered T1 and captured T2 images) is an assumption rather than a fitted parameter, and the paper does not define ground-truth change areas in terms of that rule, so the absence of precision/recall evaluation is a correctness risk, not circularity. GIM, DUSt3R, and RealityCapture are independent published tools. The prior-path objective (Eq. 5) is a coverage/redundancy optimization inspired by Zhou et al. but with its own changeability-based redundancy term; no weight (omega=3, gamma=2, phi=0.3, K=10) is fitted to the reported IoU or flight-time numbers. The real-time path (Eqs. 8-10) selects views maximizing visibility-weighted changeability of the detected target and is not equivalent to the output by construction. The efficiency comparison in Table 3 directly reflects focusing reconstruction paths on detected hulls, but the contribution is detecting those hulls, and the baseline is an external planner rather than a quantity defined by the method itself. Self-citations (UrbanBIS, Zhou et al., Liu et al.) are to published artifacts used as datasets or baselines and are not load-bearing in the derivation. No circular step was found.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method depends on hand-set parameters and modeling assumptions about priors and change detection, but the WUSU statistics are external data rather than quantities fitted to the test scenes. No new physical entities are introduced.

free parameters (5)

omega and gamma in Eq. (4) = omega = 3, gamma = 2
Hand-set scaling parameters controlling the contribution of unvisited and visited views; no sensitivity study is provided.
IoU threshold phi = phi = 0.3
Used in Sec. 5 to decide whether a new point cloud replaces or merges with the current change target.
Top-K candidate views = K = 10
Used in Sec. 6 to prevent extreme trajectory deviations; the value is stated without ablation.
Safe flight altitude h = h = 120 meters
Fixed flight altitude used in all experiments; the paper acknowledges it causes problems for variable building heights.
Poisson disk sampling radius = 15 meters for prior path, 5 meters for real-time path
Supplementary Sec. A.1 sets these radii to balance broad coverage versus detailed change exploration.

assumptions (5)

domain assumption Lack of feature matches between rendered T1 and captured T2 images indicates geometric change.
Used in Sec. 5 to define Idiff_j and mask the point cloud reconstruction; false positives and false negatives are not modeled explicitly.
domain assumption The T1 reconstruction is accurate enough that a synthetic image at the same pose is comparable to a real T2 image.
Sec. 5 compares I_T2_j with I_T1_j rendered from the T1 model; holes or artifacts in T1 would be read as changes.
domain assumption Global per-class change probabilities from WUSU transfer to UrbanBIS scenes.
Sec. 3.2 and Table 4 assign q(s_i) = p_i based on WUSU class change rates; there is no validation that these priors hold in the target scenes.
domain assumption A fixed safe height of 120 meters with five camera orientations can observe all target surfaces.
Supplementary Sec. A.1 generates all views on a safe-height plane; the paper's limitations section says tall and short buildings are handled poorly.
standard math Visibility, Poisson disk sampling, and TSP solvers behave as standard tools.
Used in Sec. 3.1, Sec. 4, and Supplementary Sec. B.2 without proof; standard background.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Aerial Path Online Planning for Urban Scene Updation." pith.science (2026). https://pith.science/paper/ZZ3HSHNS

@misc{pith2026250501486,
  author       = {Pith},
  title        = {Pith review of: Aerial Path Online Planning for Urban Scene Updation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZ3HSHNS}},
  note         = {Machine review of arXiv:2505.01486}
}

read the original abstract

We present the first scene-update aerial path planning algorithm specifically designed for detecting and updating change areas in urban environments. While existing methods for large-scale 3D urban scene reconstruction focus on achieving high accuracy and completeness, they are inefficient for scenarios requiring periodic updates, as they often re-explore and reconstruct entire scenes, wasting significant time and resources on unchanged areas. To address this limitation, our method leverages prior reconstructions and change probability statistics to guide UAVs in detecting and focusing on areas likely to have changed. Our approach introduces a novel changeability heuristic to evaluate the likelihood of changes, driving the planning of two flight paths: a prior path informed by static priors and a dynamic real-time path that adapts to newly detected changes. The framework integrates surface sampling and candidate view generation strategies, ensuring efficient coverage of change areas with minimal redundancy. Extensive experiments on real-world urban datasets demonstrate that our method significantly reduces flight time and computational overhead, while maintaining high-quality updates comparable to full-scene re-exploration and reconstruction. These contributions pave the way for efficient, scalable, and adaptive UAV-based scene updates in complex urban environments.

Figures

Figures reproduced from arXiv: 2505.01486 by the authors.

**Figure 1.** We present the first aerial trajectory planning method for urban scene updates. Starting with a previously explored scene ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** The input to our method is a labeled reconstructed model of [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** An example of our real-time path (detailed in Sec. 6) applied to a scene from UrbanBIS [Yang et al [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: Visual comparison of the efficiency of our method (b) versus Zhou et al [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Qualitative comparison of path planning methods for detecting [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

**Figure 6.** Figure 6: Visual comparison of reconstructed buildings using either the [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Qualitative results for two large-scale scenes from the UrbanBIS [Yang et al [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: We test our method on one real-world scene where [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]

**Figure 9.** Figure 9: (left) An illustration of our candidate view generation strategy. (right) [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: Results on a real-world scene featuring newly constructed buildings and updated areas. Our method successfully detects all change areas, though [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

**Figure 11.** Figure 11: Failure case where the ground truth (GT) change area is a single region, but our method incorrectly identifies it as two separate convex hulls. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]

**Figure 12.** Figure 12: Visual comparison of reconstructed buildings using either the [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 4 canonical work pages

[2]

next-best-view

Receding horizon" next-best-view" planner for 3d exploration. In 2016 IEEE international conference on robotics and automation (ICRA) . IEEE, 1462–1468. Titus Cieslewski, Elia Kaufmann, and Davide Scaramuzza. 2017. Rapid exploration with multi-rotors: A frontier selection method for high speed flight. In 2017 IEEE/RSJ International Conference on Intellige...

arXiv 2012
[2016]

In Computer vision– ECCV 2016 workshops: Amsterdam, the Netherlands, October 8-10 and 15-16, 2016, proceedings, part II 14

Fully-convolutional siamese networks for object tracking. In Computer vision– ECCV 2016 workshops: Amsterdam, the Netherlands, October 8-10 and 15-16, 2016, proceedings, part II 14 . Springer, 850–865. Andreas Bircher, Mina Kamel, Kostas Alexis, Helen Oleynikova, and Roland Siegwart

work page 2016
[2018]

In Proceedings of the European conference on computer vision (ECCV) workshops

ChangeNet: A deep learning architecture for visual change detection. In Proceedings of the European conference on computer vision (ECCV) workshops . 0–0. Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud

work page
[2020]

ACM Transactions on Graphics (TOG) 39, 6 (2020), 1–16

Offsite aerial path planning for efficient urban scene reconstruction. ACM Transactions on Graphics (TOG) 39, 6 (2020), 1–16. 11 SIGGRAPH ’25 Conference Proceedings, August 10–14, 2025, Vancouver, BC, Canada M. Tang, N. Wang, Z. Xie, J. Hu, K. Xie, X. Guo, and H. Huang SUPPLEMENTARY MATERIAL This section provides additional details and experimental result...

work page 2020
[2024]

In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition . 20697–20709. Shihao Wu, Wei Sun, Pinxin Long, Hui Huang, Daniel Cohen-Or, Minglun Gong, Oliver Deussen, and Baoquan Chen. 2014. Quality-driven poisson-guided autoscanning. (2014). Kai Xu, Lintao Zheng, Zihao Yan, Guohang Yan, Eug...

work page 2014

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [2]

next-best-view

Receding horizon" next-best-view" planner for 3d exploration. In 2016 IEEE international conference on robotics and automation (ICRA) . IEEE, 1462–1468. Titus Cieslewski, Elia Kaufmann, and Davide Scaramuzza. 2017. Rapid exploration with multi-rotors: A frontier selection method for high speed flight. In 2017 IEEE/RSJ International Conference on Intellige...

arXiv 2012

[2] [2016]

In Computer vision– ECCV 2016 workshops: Amsterdam, the Netherlands, October 8-10 and 15-16, 2016, proceedings, part II 14

Fully-convolutional siamese networks for object tracking. In Computer vision– ECCV 2016 workshops: Amsterdam, the Netherlands, October 8-10 and 15-16, 2016, proceedings, part II 14 . Springer, 850–865. Andreas Bircher, Mina Kamel, Kostas Alexis, Helen Oleynikova, and Roland Siegwart

work page 2016

[3] [2018]

In Proceedings of the European conference on computer vision (ECCV) workshops

ChangeNet: A deep learning architecture for visual change detection. In Proceedings of the European conference on computer vision (ECCV) workshops . 0–0. Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud

work page

[4] [2020]

ACM Transactions on Graphics (TOG) 39, 6 (2020), 1–16

Offsite aerial path planning for efficient urban scene reconstruction. ACM Transactions on Graphics (TOG) 39, 6 (2020), 1–16. 11 SIGGRAPH ’25 Conference Proceedings, August 10–14, 2025, Vancouver, BC, Canada M. Tang, N. Wang, Z. Xie, J. Hu, K. Xie, X. Guo, and H. Huang SUPPLEMENTARY MATERIAL This section provides additional details and experimental result...

work page 2020

[5] [2024]

In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition . 20697–20709. Shihao Wu, Wei Sun, Pinxin Long, Hui Huang, Daniel Cohen-Or, Minglun Gong, Oliver Deussen, and Baoquan Chen. 2014. Quality-driven poisson-guided autoscanning. (2014). Kai Xu, Lintao Zheng, Zihao Yan, Guohang Yan, Eug...

work page 2014