Pith. sign in

REVIEW 3 major objections 4 minor 29 references

Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes

T0 review · 3 major / 4 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 This paper claims that interactive 3D point-cloud segmentation becomes more click-efficient and uncertainty-aware when framed as hierarchical neural-process prediction, and that the resulting model NPISeg3D beats AGILE3D by 4.2 to 8.4…

desk verdict A credible first NP-based interactive 3D segmentation paper with real accuracy gains; the uncertainty headline needs quantitative support. read the letter →

arxiv 2505.01726 v2 pith:PB3HFQ56 submitted 2025-05-03 cs.CV

Jie Liu , Pan Zhou , Zehao Xiao , Jiayi Shen , Wenzhe Yin , Jan-Jakob Sonke , Efstratios Gavves This is my paper · ORCID
classification cs.CV
keywords interactive3Dsegmentationneuralprocesseshierarchicallatentvariablesuncertaintyestimationfew-shotgeneralizationpointcloudclick-basedannotationprobabilisticprototypemodulator
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

The paper argues that interactive 3D point-cloud segmentation — a user clicks points in a 3D scene to obtain object masks — should be treated as a probabilistic learning problem rather than a deterministic click-conditioned mapping, because only a probabilistic model can both generalize from very few clicks and say where it is unsure. It builds NPISeg3D on neural processes: the clicks form a context set, the unlabeled scene points form a target set, and the model outputs a distribution over masks. A hierarchy of latent variables, one for the whole scene and one per object, is meant to capture global layout together with click-specific object detail, and a probabilistic prototype modulator turns every click prototype into a stochastic classifier that can be sampled. On four datasets the framework reports higher IoU with fewer clicks than prior baselines, for instance 4.2 and 8.4 mean-IoU points over AGILE3D on the out-of-domain KITTI-360 in multi- and single-object segmentation, and it outputs variance-based uncertainty maps that concentrate on boundaries and misclassified regions. If the claims hold, neural processes are a valid route to both few-shot generalization and usable confidence estimates for click-based 3D annotation.

What carries the argument

The load-bearing machinery is the pair of hierarchical latent variables — scene-level $z_s$ and object-level $z_o^m$ — plus the probabilistic prototype modulator of Eq. (8). The modulator is a scale-and-shift transform, $\tilde{X}_C^{m,i,j} = \gamma(z_o^{m,j}) \odot X_C^{m,i} + \beta(z_o^{m,j})$, in which a two-layer MLP maps each Monte Carlo sample of the object latent to feature-wise parameters; this is what converts the deterministic click prototypes of earlier attention-based models into stochastic, object-aware classifiers. The formal spine is the factorized predictive distribution of Eq. (5), in which objects are scored independently given the shared scene latent; from it follow the factorized ELBO of Eq. (9) with its two KL regularizers and the per-point max rule of Eq. (11) that assigns each point to the most confident object. Scene- and object-level aggregators — a single-layer transformer over averaged prototypes and an MLP over a convex combination of $z_s$ with click features — parameterize the prior and posterior Gaussians, forming the information flow from scene to objects to clicks.

What would settle it

Build an evaluation split by object-overlap ratio: on scenes with heavy overlap or occlusion, the per-point max over independently scored objects should misassign boundary points and the predicted uncertainty should concentrate on the wrong side of occlusion boundaries, while on isolated well-separated objects the gap to a joint model should shrink — a pattern that would confirm the independence assumption is doing the work. As a cheaper second check, the paper shows uncertainty maps but never calibrates them against actual errors, so measuring whether its variance maps predict per-point correctness (AUROC or reliability curves) better than MC dropout at matched compute would settle whether the uncertainty claim is stronger than a standard alternative.

Watch

Extended reading notes

Core claim

NPISeg3D's central claim is that interactive 3D segmentation can be modeled as the predictive distribution $p(Y_T \mid X_T, D_C)$ of a neural process, with the user's clicks (features plus one-hot labels) as the context set and all unclicked scene points as targets. A hierarchical latent structure handles multi-object scenes: a scene-level latent $z_s$ inferred from all clicks encodes global layout and inter-object relations, and per-object latents $z_o^m$, conditioned on $z_s$ and the object's own click prototypes, encode object-specific detail. A probabilistic prototype modulator then applies a feature-wise affine transform $\tilde{X}_C^{m,i,j} = \gamma(z_o^{m,j}) \odot X_C^{m,i} + \beta(z_o^{m,j})$ to each click prototype, with scale and shift produced by a two-layer MLP from Monte Carlo samples of $z_o^m$, so each click acts as a sampled classifier rather than a fixed vector. Training maximizes the ELBO of Eq. (9), in which priors inferred from clicks alone are pulled toward variational posteriors inferred from full target masks by two KL-divergence terms; at test time, mask logits average cosine similarity over latent samples and take a per-object max over clicks (Eq. (11)), and replacing the average with the variance yields the uncertainty map. The paper claims this is the first probabilistic framework for interactive 3D segmentation and reports superior few-shot generalization with reliable uncertainty across ScanNet, S3DIS, Replica, and KITTI-360.

Load-bearing premise

The model assumes each object's segmentation function is conditionally independent of every other object once the shared scene latent and that object's own latent are fixed; if objects overlap, occlude one another, or receive correlated clicks, the per-object factorization behind Eqs. (9) and (11) will misassign boundary points, so the reported gains depend on learned features staying separable enough for the per-point max rule to win.

Editorial extensions

If this is right

  • Interactive segmentation becomes a distribution over masks rather than a single deterministic output: the same clicks can be sampled many times, which is exactly what makes the uncertainty maps possible.
  • Domain transfer improves without retraining: on the out-of-domain KITTI-360, mean IoU rises by 4.2 points (multi-object) and 8.4 points (single-object) over AGILE3D, and the number of clicks needed to reach 80, 85, and 90 percent IoU drops.
  • Uncertainty concentrates on object boundaries, misclassified regions, and points far from any click, and it shrinks as clicks accumulate, so an annotation interface can point users to the next most useful click; a user study with real human annotators reaches simulator-level performance.
  • The probabilistic treatment is more stable than MC dropout for confidence estimation: NPISeg3D's performance is flat across 5, 10, and 20 latent samples, while MC dropout degrades as the dropout rate grows.
  • The framework extends to fine-grained part segmentation on PartNet and benefits from feeding the previous mask back into the model, consistent with the claim that hierarchical context, not just click detail, drives the gains.

Reading between the lines

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

  • Nothing in the formulation restricts it to point clouds: a context set of clicks, hierarchical latents, and a probabilistic modulator could equally apply to 2D interactive segmentation or volumetric medical images, where uncertainty-guided clicking is already a recognized need; the paper does not make this claim.
  • The uncertainty map in Eq. (13) measures disagreement among the sampled click classifiers, not the probability of error, and the paper reports no calibration curve or AUROC against actual mistakes; a direct extension would test whether the variance predicts the IoU gain of the next click better than random click selection.
  • The independence assumption marks a testable boundary — overlap and occlusion should expose it. A split of the benchmark by object-overlap ratio, or a synthetic stacking scenario, would show where the per-point max starts failing and whether a joint association layer would close the gap.
  • The ablations hint the scene-level latent earns its keep unevenly — little effect on S3DIS, more on KITTI-360 — so a cheap follow-up hypothesis, that the optimal hierarchy depth tracks scene complexity and object density, is testable by scaling the number of interacting objects per scene.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes NPISeg3D, a hierarchical neural-process (NP) framework for interactive 3D point cloud segmentation. User-provided clicks are treated as a context set, and the model infers a scene-level latent variable z_s and per-object latent variables z_o through variational inference. A probabilistic prototype modulator rescales click prototypes from samples of z_o, and inference uses Monte Carlo aggregation with a per-point max rule. Experiments on ScanNet, S3DIS, Replica, KITTI-360, and PartNet compare against InterObject3D, InterObject3D++, and AGILE3D, reporting IoU and NoC metrics, ablations over components and uncertainty strategies, and a user study with 10 participants.

Significance. If the segmentation results are confirmed, the paper makes a useful contribution: it introduces the first NP-based formulation for click-based interactive 3D segmentation and demonstrates consistent, sometimes large, out-of-domain gains over AGILE3D (e.g., Table 2 KITTI-360: IoU@5 55.7 vs. 44.4). The ELBO derivation in Appendix A is a correct application of variational inference, and the implementation details are largely transparent, including training hyperparameters and computational cost comparisons. However, the paper's second headline contribution, 'reliable uncertainty estimation,' is asserted in the abstract and conclusion but is not supported by any quantitative uncertainty-quality evaluation; the current evidence is qualitative or uses accuracy metrics that do not measure uncertainty. The in-domain accuracy gains are also small and are reported without error bars.

major comments (3)
  1. [Section 5.3, Appendix B.5 (Eqs. 13-14), Table 7] The central claim that NPISeg3D 'provides reliable uncertainty estimations' is not supported by any quantitative evaluation of uncertainty quality. Table 6 reports mIoU@10 and NoC@85 for MC Dropout versus NPISeg3D, but these are segmentation accuracy and click-efficiency metrics; they do not measure calibration, discrimination, or the correlation between the uncertainty map and per-point errors. Figures 4, 7, and 8 are qualitative, and the user study in Table 7 has no control condition without the uncertainty overlay, so the causal effect of uncertainty on annotation efficiency is untested. Please add (a) standard uncertainty metrics such as expected calibration error, misclassification-detection AUROC, or per-point error correlation; (b) a comparison against a baseline uncertainty method (e.g., softmax entropy or MC Dropout) under those metrics; and (c) a user study arm without uncertainty guidance or a click-target analysis.
  2. [Tables 1-3] The main quantitative tables report single-run results without standard deviations or significance tests. In-domain margins are small (Table 1 ScanNet40: IoU@5 82.6 vs. 82.3 for AGILE3D; Table 2 ScanNet: IoU@5 80.5 vs. 79.9), so it is unclear whether the claimed superiority is statistically meaningful. Please provide means and standard deviations over multiple runs, or at least significance tests for the main comparisons.
  3. [Section 4.1 (Eq. (4)) and Eq. (11)] The model assumes object-level segmentation functions are conditionally independent given the scene and object latents, and inference uses a per-point max over independently computed object scores. For overlapping or occluding objects, this factorization can misassign boundary points. The paper neither analyzes this failure mode nor evaluates on scenes with heavy occlusion; please add such an analysis or explicitly discuss this limitation as a caveat on the multi-object claims.
minor comments (4)
  1. [Eq. (7), Appendix D.2] Appendix D.2 reports lambda_kl = 0.005 but does not report alpha in Eq. (7) or Nzo for the main results; please add these hyperparameter values.
  2. [Throughout] There are several typos, e.g., 'hierarchal' in Section 4.2, 'furhter' in Section 5.2, 'traing' in the Table 11 caption, and 'spare voxels' instead of 'sparse voxels' in Appendix D.2.
  3. [Appendix B.1] Appendix B.1 states that object ids for the PartNet subset will be released; since the subset is randomly sampled, please include the ids or a fixed seed in the paper for reproducibility.
  4. [Appendix B.5, Eq. (13)] Eq. (13) has a formatting issue: the variance expression contains an apparent misplaced comma and unbalanced parentheses.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: NPISeg3D's accuracy is benchmarked against held-out external baselines, and its uncertainty maps are computed from MC variance of the model's own predictions rather than fitted to error. The only caveat is that the 'reliable uncertainty' claim lacks quantitative validation, which is a support gap, not circularity.

full rationale

The paper's central derivation chain is self-contained. Equations (3)-(5) define a hierarchical NP generative model; Eq. (9) is a standard ELBO with variational posteriors q(z_s|X_T), q(z_o|z_s,X_T) and priors p(z_s|X_C), p(z_o|z_s,X_C). The KL terms train the priors to match posteriors, and at test the priors are evaluated on user clicks. The segmentation logits in Eq. (11) and uncertainty map in Eqs. (13)-(14) are Monte Carlo means and variances of the same cosine-similarity scores; neither is fitted to ground-truth error maps nor reused as a fitted constant. Tables 1-3 report IoU and NoC on ScanNet, S3DIS, Replica, KITTI-360, and PartNet against external baselines (InterObject3D, InterObject3D++, AGILE3D), so the accuracy claims are externally grounded. The paper cites co-author NP papers (e.g., Shen et al., 2021; 2023) for background motivation, but the contribution is validated by its own ablations (Table 4), so those citations are not load-bearing. The 'reliable uncertainty estimation' claim is supported only by qualitative visualizations (Figures 4, 7, 8) and Table 6, which reports segmentation accuracy rather than calibration or discrimination of uncertainty; this is a missing-support/correctness concern, not a circular reduction. No equation in the paper reduces to its input by construction, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The model is a conditional neural process with an added hierarchical latent: scene-level z_s and per-object z_o. Central new machinery is the probabilistic prototype modulator, which applies FiLM-style scale/shift to click prototypes from sampled object latents. No new physical entities are proposed; the latent variables are internal stochastic model components. The main free choices are the balance weight alpha (unreported), KL weight 0.005, the number of Monte Carlo samples (unreported for main tables), and the RITM iterative click count (unreported). The key domain assumption is conditional independence of objects, which drives the factorized ELBO and the max-based inference. The test-time oracle click simulation is a further assumption that the user study only partially validates.

free parameters (4)
  • alpha (alpha) = not reported; claimed in [0,1]
    Balances scene-level latent and object click context in Eq. (7); value not given in paper, so an implementation must choose it.
  • lambda_kl = 0.005
    KL regularization coefficient in Eq. (10); chosen manually, affects how strongly priors are pushed toward test-time posteriors.
  • Nzo (Monte Carlo samples) = main results not stated; ablation uses 5, 10, 20
    Number of latent samples in Eq. (11) and uncertainty map Eq. (13); the paper reports robustness across 5-20 but not which is used for main tables.
  • RITM iterative click count Niter = not reported
    Number of iterative clicks during training in the RITM strategy (Appendix D.2); not specified, affects the training distribution.
assumptions (5)
  • standard math The ELBO derivation uses standard Jensen inequality and amortized variational inference (Appendix A).
    Unproved background from Kingma and Welling; necessary for Eq. (9).
  • domain assumption Object-level segmentation functions are conditionally independent given scene latent z_s and object latents z_o (Eq. 4, Sec. 4.1).
    Justifies factorization over objects; realistic only if clicks and objects are separable and occlusion effects are weak.
  • domain assumption Cosine similarity in the learned feature space is a sufficient decision rule for assigning points to objects (Eq. 11).
    The mask head is implemented as max over click prototypes of averaged cosine; this presupposes linearly separable, well-clustered features.
  • domain assumption Test-time click simulation using ground-truth misclassification regions approximates human annotator behavior (Appendix D.2).
    Standard benchmark assumption in interactive segmentation; user study partially checks it but with no competing method.
  • domain assumption The point encoder (Minkowski Res16UNet34C) and attention network provide features that transfer across sensors and domains (Appendix D.2).
    Out-of-domain generalization claim depends on feature transferability; not proven by construction.
invented entities (2)
  • Scene-specific latent variable z_s
    purpose: Captures global scene context and inter-object relationships (Eq. 5-6).
    Internal stochastic variable introduced by the model; no external falsifiable handle beyond the model's predictive performance.
  • Object-specific latent variable z_o
    purpose: Captures object-level characteristics and drives probabilistic prototype modulation (Eq. 7-8).
    Internal stochastic variable; no independent evidence outside the system.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes." pith.science (2026). https://pith.science/paper/PB3HFQ56

@misc{pith2026250501726,
  author       = {Pith},
  title        = {Pith review of: Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PB3HFQ56}},
  note         = {Machine review of arXiv:2505.01726}
}
read the original abstract

Interactive 3D segmentation has emerged as a promising solution for generating accurate object masks in complex 3D scenes by incorporating user-provided clicks. However, two critical challenges remain underexplored: (1) effectively generalizing from sparse user clicks to produce accurate segmentation, and (2) quantifying predictive uncertainty to help users identify unreliable regions. In this work, we propose NPISeg3D, a novel probabilistic framework that builds upon Neural Processes (NPs) to address these challenges. Specifically, NPISeg3D introduces a hierarchical latent variable structure with scene-specific and object-specific latent variables to enhance few-shot generalization by capturing both global context and object-specific characteristics. Additionally, we design a probabilistic prototype modulator that adaptively modulates click prototypes with object-specific latent variables, improving the model's ability to capture object-aware context and quantify predictive uncertainty. Experiments on four 3D point cloud datasets demonstrate that NPISeg3D achieves superior segmentation performance with fewer clicks while providing reliable uncertainty estimations.

Figures

Figures reproduced from arXiv: 2505.01726 by the authors.

Figure 1
Figure 1. Overview of NPISeg3D. We formulate interactive 3D segmentation as a probabilistic modeling problem with neural processes. Given a 3D scene S and a user click set C, a point encoder encodes them into click prototypes XC (context data) and scene features XT (target data). Then, we introduce two hierarchical latent variables: scene-level latent variable zs and object-level latent variable zo, to enable probabilistic mo… view at source ↗
Figure 2
Figure 2. Computational Graph of our NPISeg3D. The frame￾work incorporates a hierarchical inference structure with a scene￾level latent variable (zs) and an object-level latent variable (zo), capturing contextual information at different spatial levels. 4.2. Hierarchical NP Framework for segmentation To enhance few-shot generalization in complex scenes, we introduce a hierarchical latent variable structure into the NP framewo… view at source ↗
Figure 3
Figure 3. Qualitative results on interactive multi-object segmen￾tation on S3DIS and KITTI360. Newly added clicks are repre￾sented by dark-colored dots. Please zoom in for more details. and KITTI-360 (Liao et al., 2022), an outdoor LiDAR dataset designed for autonomous driving scenarios. Further details on the dataset setup are provided in Appendix D.1. Evaluation Metrics. We follow prior works (Kontogianni et al., 2023; Yue … view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Uncertainty maps of predictions with increasing clicks. We show mask predictions and uncertainty after K clicks. explore alternative modulation strategies, such as Concat and Add, to assess different ways of integrating object￾specific latent variables with click proto…
Figure 5
Figure 5. Figure 5: Qualitative comparison between the state-of-the-art (SoTA) method AGILE3D and our proposed NPISeg3D on the interactive multi-object segmentation task. Newly added clicks are represented by dark-colored dots. Our NPISeg3D consistently achieves higher IoU scores with the…
Figure 6
Figure 6. Figure 6: Qualitative comparison between the state-of-the-art (SoTA) method AGILE3D and our proposed NPISeg3D on the interactive single-object segmentation task. Newly added clicks are represented by dark-colored dots. Our NPISeg3D consistently achieves higher IoU scores with th…
Figure 9
Figure 9. Figure 9: User interface for interactive segmentation with integrated uncertainty estimation as guidance. In the main window, users provide input clicks, and the model generates a segmentation prediction. The upper-right window displays uncertainty estimation for the current seg…
Figure 7
Figure 7. Figure 7: Qualitative Results on Interactive Single-Object Segmentation. In the multi-object segmentation task, IoU@k denotes the average Intersection over Union (IoU) after k clicks per object. For clarity, only the first click per object is visualized, with subsequent clicks o…
Figure 8
Figure 8. Figure 8: More qualitative results on the interactive single-object segmentation task. In multi-object segmentation task, IoU@k denotes the average IoU after k clicks per object. We show the first click per object, and omit following clicks for better visualization. In the uncer…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 15 canonical work pages

  1. [3]

    Deep gaussian processes: A survey

    Jakkala, K. Deep gaussian processes: A survey. arXiv preprint arXiv:2106.12135,

    arXiv

  2. [9]

    Interactive object segmentation in 3d point clouds

    Kontogianni, T., Celikkan, E., Tang, S., and Schindler, K. Interactive object segmentation in 3d point clouds. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 2891–2897. IEEE,

    work page 2023

  3. [10]

    Uncertainty quantification with statistical guarantees in end-to-end autonomous driv- ing control

    Michelmore, R., Wicker, M., Laurenti, L., Cardelli, L., Gal, Y ., and Kwiatkowska, M. Uncertainty quantification with statistical guarantees in end-to-end autonomous driv- ing control. In 2020 IEEE international conference on robotics and automation (ICRA), pp. 7344–7350. IEEE,

    work page 2020

  4. [11]

    and Grover, A

    Nguyen, T. and Grover, A. Transformer neural processes: Uncertainty-aware meta learning via sequence modeling. arXiv preprint arXiv:2207.04179,

    arXiv

  5. [12]

    Mask3d: Mask transformer for 3d semantic instance segmentation

    Schult, J., Engelmann, F., Hermans, A., Litany, O., Tang, S., and Leibe, B. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE Interna- tional Conference on Robotics and Automation (ICRA), pp. 8216–8223. IEEE,

    work page 2023

  6. [13]

    H., and Knoll, A

    Shafaei, S., Kugele, S., Osman, M. H., and Knoll, A. Un- certainty in machine learning: A safety perspective on autonomous driving. In Computer Safety, Reliability, and Security: SAFECOMP 2018 Workshops, ASSURE, DECSoS, SASSUR, STRIVE, and WAISE, V¨aster˚as, Swe- den, September 18, 2018, Proceedings 37, pp. 458–464. Springer,

    work page 2018

  7. [15]

    Interactive an- notation of 3d object geometry using 2d scribbles

    Shen, T., Gao, J., Kar, A., and Fidler, S. Interactive an- notation of 3d object geometry using 2d scribbles. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVII 16, pp. 751–767. Springer,

    work page 2020

  8. [16]

    A., and Konushin, A

    Sofiiuk, K., Petrov, I. A., and Konushin, A. Reviving itera- tive training with mask guidance for interactive segmen- tation. In 2022 IEEE International Conference on Image Processing (ICIP), pp. 3141–3145. IEEE,

    work page 2022

Show all 29 references
  1. [17]

    J., Mur-Artal, R., Ren, C., Verma, S., et al

    Straub, J., Whelan, T., Ma, L., Chen, Y ., Wijmans, E., Green, S., Engel, J. J., Mur-Artal, R., Ren, C., Verma, S., et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797,

    1906 arXiv

  2. [18]

    W., Pollefeys, M., Tombari, F., and Engelmann, F

    Takmaz, A., Fedele, E., Sumner, R. W., Pollefeys, M., Tombari, F., and Engelmann, F. Openmask3d: Open- vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631,

    arXiv

  3. [19]

    Bridge the infer- ence gaps of neural processes via expectation maximiza- tion

    Wang, Q., Federici, M., and van Hoof, H. Bridge the infer- ence gaps of neural processes via expectation maximiza- tion. arXiv preprint arXiv:2501.03264,

    arXiv

  4. [21]

    Agile3d: Attention guided interactive multi-object 3d segmentation

    Yue, Y ., Mahadevan, S., Schult, J., Engelmann, F., Leibe, B., Schindler, K., and Kontogianni, T. Agile3d: Attention guided interactive multi-object 3d segmentation. arXiv preprint arXiv:2306.00977,

    arXiv

  5. [22]

    Refin- ing segmentation on-the-fly: An interactive framework for point cloud semantic segmentation

    Zhang, P., Wu, T., Sun, J., Li, W., and Su, Z. Refin- ing segmentation on-the-fly: An interactive framework for point cloud semantic segmentation. arXiv preprint arXiv:2403.06401,

  6. [23]

    Y ., Xiang, F., and Su, H

    Zhou, Y ., Gu, J., Chiang, T. Y ., Xiang, F., and Su, H. Point- sam: Promptable 3d segmentation model for point clouds. arXiv preprint arXiv:2406.17741,

    arXiv

  7. [24]

    However, iterative training incurs high computational costs, making it less practical for large-scale applications

    demonstrated the effectiveness of iterative training for interactive multi-object segmentation. However, iterative training incurs high computational costs, making it less practical for large-scale applications. In contrast, we adopt the more efficient RITM sampling strategy (...

    2022

  8. [25]

    and train our model on the ScanNetV2-Train dataset (Dai et al., 2017), which contains 1,200 indoor scenes. Evaluation is conducted across four diverse datasets, covering both indoor and outdoor environments: ScanNetV2-Val (Dai et al., 2017), S3DIS (Armeni et al., 2016), Replic...

    2017

  9. [26]

    KITTI-360 (Liao et al., 2022), an outdoor LiDAR point cloud dataset, is designed for 3D perception tasks in autonomous driving and provides large-scale outdoor scenes

    is a photorealistic indoor dataset comprising high-quality 3D reconstructions of real-world environments, offering diverse scene layouts and materials for testing model generalization. KITTI-360 (Liao et al., 2022), an outdoor LiDAR point cloud dataset, is designed for 3D perc...

    2022

  10. [27]

    backbone, following prior works (Kontogianni et al., 2023; Yue et al., 2023; Schult et al., 2023). The 3D scene is first quantized intoN ′ spare voxels with a fixed resolution of 5cm, ensuring efficient and consistent representation, as done in previous studies (Kontogianni et...

    2023

  11. [28]

    This iterative approach closely mimics the test-time click sampling strategy, enhancing the model’s ability to refine segmentation with progressive user feedback

    has been demonstrated to effectively improve segmentation performance by simulating user clicks iteratively based on the model’s predictions from previous rounds. This iterative approach closely mimics the test-time click sampling strategy, enhancing the model’s ability to ref...

    2022

  12. [29]

    The high computational cost makes it impractical for large-scale training scenarios. In contrast, interactive 2D segmentation commonly employs a more efficient strategy, known 19 Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes as Random and Iterati...

    2022

  13. [2013]

    Con- tinuous adaptation for interactive object segmentation by learning from corrections

    Kontogianni, T., Gygli, M., Uijlings, J., and Ferrari, V . Con- tinuous adaptation for interactive object segmentation by learning from corrections. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVI 16 , pp. 579–596....

    2020

  14. [2018]

    Multi-task neural processes

    Shen, J., Zhen, X., Worring, M., and Shao, L. Multi-task neural processes. arXiv preprint arXiv:2111.05820,

    arXiv

  15. [2019]

    Neural processes with stochastic attention: Paying more attention to the context dataset

    Kim, M., Go, K., and Yun, S.-Y . Neural processes with stochastic attention: Paying more attention to the context dataset. arXiv preprint arXiv:2204.05449,

    arXiv

  16. [2020]

    W., Rezende, D., and Eslami, S

    Garnelo, M., Rosenbaum, D., Maddison, C., Ramalho, T., Saxton, D., Shanahan, M., Teh, Y . W., Rezende, D., and Eslami, S. A. Conditional neural processes. In Interna- tional conference on machine learning, pp. 1704–1713. PMLR, 2018a. Garnelo, M., Schwarz, J., Rosenbaum, D., Vi...

    arXiv

  17. [2021]

    E., and Yao, L

    Jha, S., Gong, D., Wang, X., Turner, R. E., and Yao, L. The neural process family: Survey, applications and perspec- tives. arXiv preprint arXiv:2209.00517,

    arXiv

  18. [2022]

    9 Probabilistic Interactive 3D Segmentation with Hierarchical Neural Processes Kingma, D. P. and Welling, M. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,

    arXiv

  19. [2023]

    Versatile neural processes for learning implicit neural representa- tions

    Guo, Z., Lan, C., Zhang, Z., Lu, Y ., and Chen, Z. Versatile neural processes for learning implicit neural representa- tions. arXiv preprint arXiv:2301.08883,

    arXiv

  20. [2024]

    Kim, H., Mnih, A., Schwarz, J., Garnelo, M., Eslami, A., Rosenbaum, D., Vinyals, O., and Teh, Y . W. Attentive neural processes. arXiv preprint arXiv:1901.05761,

    1901 arXiv

  21. [2025]

    Wu, J., Zhang, C., Li, Z., Fu, H., Peng, X., and Zhou, J. T. dugmatting: decomposed-uncertainty-guided mat- ting. arXiv preprint arXiv:2306.01452,

    arXiv

Pith tools

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