Pith. sign in

REVIEW 3 major objections 5 minor 51 references

Training-Free Open-Vocabulary 3D Point-Cloud Segmentation on the Generalized Few-Shot Benchmark

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A fully training-free, support-free pipeline — a frozen dense 3D vision-language prior reconciled with a frozen promptable concept segmenter by one cross-view agreement threshold — improves novel-class mIoU by +2.6 on ScanNet200 and +15.7 o

desk verdict A genuinely training-free floor for GFS-PCS plus an honest support-free negative result, but the headline gains rest on a per-benchmark threshold chosen with evaluation data in view—and the promised defense is in a missing supplementary. read the letter →

arxiv 2607.15331 v1 pith:UBCLX2KZ submitted 2026-07-16 cs.CV

Silas kwabla Gah , Ebenezer Owusu This is my paper · ORCID
classification cs.CV
keywords open-vocabulary3Dsegmentationgeneralizedfew-shotlearningtraining-freecross-viewconsistencypoint-cloudsceneunderstandingmulti-viewfusion
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

This paper tries to establish how much of generalized few-shot 3D point-cloud segmentation can be done with none of the usual supervision: no training, no 3D labels, and not even the few-shot support examples. Its claim is that a frozen dense 3D vision-language model and a frozen promptable 2D segmenter, prompted only by novel class names and joined by a single cross-view agreement threshold, improve novel-class accuracy by 2.6 mIoU on ScanNet200 and 15.7 mIoU on ScanNet++ over the training-free dense prior, while keeping base accuracy nearly intact. On ScanNet200 this recovers a third of the novel-class gap to the trained state of the art; on ScanNet++ it nearly doubles novel-class accuracy and lifts the base-novel harmonic mean from 21.5 to 31.1, within 0.7 of the trained method. A reader should care because the result suggests the expensive training and annotation pipeline in this benchmark may be partially replaceable by a parameter-free consensus rule, and because the paper goes on to argue that the few-shot support adds no benefit once consensus is applied.

What carries the argument

The load-bearing mechanism is the cross-view agreement score s_i(n): for each point i and novel class n, the fraction of posed RGB frames that both see the point and mark it as n, with the winning novel class assigned when s_i = max_n s_i(n) clears a single threshold τ. It is parameter-free beyond τ, and it does two jobs at once: it encodes both how well a point is seen and how confidently the sparse segmenter names it, which is why separate visibility and confidence gates are redundant. The two frozen branches supply the inputs — a dense per-point logit field from a 3D vision-language model and per-frame concept masks from a promptable 2D segmenter lifted by projection — but the consensus s

What would settle it

On ScanNet++, fix τ using only a randomly held-out 20% of the validation scenes and then evaluate on the remaining 80%; if the novel-class gain over the dense prior falls below the reported +15.7 or becomes statistically indistinguishable from zero, the central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that cross-view consistency is the full reconciliation rule between a dense but noisy 3D vision-language prior and a precise but over-firing 2D concept segmenter. The dense branch labels every point from a frozen RegionPLC model; the sparse branch runs a promptable segmenter (SAM3) on posed RGB frames with the bare novel class names, projects the resulting masks into the point cloud, and computes, for each point and each novel class, the fraction of visible views that assign that class. A point is relabeled novel only when this fraction reaches a single per-benchmark threshold. This suppresses the single-view false positives that make naive sparse-o

Load-bearing premise

The reported gains depend on a single cross-view agreement threshold chosen per benchmark without a fully independent development split — on ScanNet200 after inspecting the full validation sweep, and on ScanNet++ from one probe scene — so honest held-out selection could shrink the gains.

Editorial extensions

If this is right

  • On ScanNet200, a fully training-free and support-free pipeline raises novel-class mIoU by +2.6 over the frozen dense prior, with base mIoU within 0.5 of the dense prior, recovering a third of the gap to the trained state of the art.
  • On ScanNet++, the same pipeline raises novel-class mIoU by +15.7 (from 16.2 to 31.9) at a 1.7 base cost, lifting the base-novel harmonic mean from 21.5 to 31.1, within 0.7 of the trained reference.
  • Cross-view consistency, not extra gates or support, is the operative mechanism: naive sparse-over-dense overwrite trades base accuracy for novel, while consensus raises both, and adding view-count visibility or dense-confidence gates leaves the harmonic mean below consistency alone.
  • Injecting the few-shot support at the two loci the frozen pipeline admits — as a fusion gate and as a prototypical dense classifier — gives no improvement (the gate is null; the classifier is significantly negative), so the method needs no support annotations.
  • The gains concentrate on small, text-promptable novel objects the dense prior misses (blinds, windowsill, telephone, pillow, keyboard, picture), while a few large classes already well segmented by the dense prior suffer small losses, explaining the small base cost.

Reading between the lines

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

  • If the dense prior is replaced by a stronger frozen 3D vision-language model, the same consensus rule should shift the floor upward; a testable prediction is that the absolute gain from consistency shrinks as the dense prior's novel-class recall improves, rather than disappearing.
  • Because the reported τ is calibrated per benchmark to view density, a view-density-normalized agreement score could make the threshold portable across captures, potentially removing the last per-benchmark hyper-parameter — an extension the paper leaves open.
  • The support-injection negative result is scoped to text-conditioned, frozen features; a model that routes support through an exemplar-conditioned 2D segmenter might still benefit, but that requires relaxing the paper's training-free boundary.
  • The benchmark's support annotations may be less informative than the protocol assumes, at least for object-like classes; this suggests comparing training-free, support-free floors as a standard reference when evaluating any new GFS-PCS method.
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 / 5 minor

Summary. The paper proposes a fully training-free, open-vocabulary pipeline for generalized few-shot 3D point-cloud segmentation. A frozen RegionPLC model supplies a dense per-point prior; a frozen SAM3 promptable segmenter, prompted with the bare novel-class names, produces per-frame 2D masks that are lifted to the point cloud via the posed RGB frames. A single cross-view-consistency threshold τ decides when the sparse novel labels overwrite the dense prediction. On ScanNet200 the method improves novel mIoU by +2.6 over the dense-only prior at a 0.5 base cost, and on ScanNet++ by +15.7 at a 1.7 base cost. The authors report bootstrap CIs for these gains and ablations showing that visibility/confidence gates and two few-shot support-injection mechanisms provide no benefit over consistency alone.

Significance. If the reported results survive a clean evaluation protocol, this is a useful contribution. The cross-view consistency rule is simple, intuitive, and shown to suppress single-view false positives. The paper is careful to compare against the training-free dense prior and to provide paired bootstrap CIs for the headline gains. The negative result on support injection is an honest and informative finding: within this frozen pipeline, the few-shot support is null through a fusion gate and harmful through a prototypical classifier. The paper also establishes a new reference point on ScanNet++ by evaluating the RegionPLC zero-shot floor. However, the central quantitative claims rest on a per-benchmark threshold whose selection is not fully independent of the evaluation data and whose defense is delegated to a supplementary file that is not part of the submitted manuscript. This must be resolved before the reported gains can be taken at face value.

major comments (3)
  1. [§5, Tables 1–3] The headline novel-mIoU gains (+2.6 and +15.7) are controlled by the threshold τ, yet the selection protocol is not present in the manuscript. The text states that τ is 'pre-committed' and 'never tuned to maximise the reported metric', but on ScanNet200 the chosen value is described as 'below the sweep optimum' after the full validation sweep of Table 3(b); on ScanNet++ τ=0.15 is 'fixed on a single probe scene' from the evaluation set, with the paper itself noting there is no separate development split. The argument that this procedure cannot inflate the result is explicitly deferred to a supplementary section that is absent from the arXiv submission. Because τ trades novel gain against base loss, and because the ScanNet++ gain is an order of magnitude larger than the ScanNet200 gain, the reader cannot rule out that the reported magnitudes are partly selection artifacts. Please either (i
  2. [§4 and §5 Implementation] The paper claims the reconciliation is 'parameter-free beyond τ' and that the pipeline has a 'single threshold'. In fact, the sparse branch uses at least two additional fixed operating points: the SAM3 concept confidence threshold of 0.3 and the frame-subsampling stride of 20 (every 20th frame). These constants affect which masks are kept and how many views vote, and hence they affect the final mIoU numbers. While these values may be reasonable, they are not swept and their influence on the headline results is not reported. Please either provide a sensitivity analysis for these two choices or qualify the claim to 'only one tuned parameter' and acknowledge the other fixed constants as part of the method.
  3. [§5 'Statistical significance'] The paired bootstrap CIs are computed over scenes for a fixed τ. They do not account for the fact that τ was selected after inspecting the full ScanNet200 validation sweep and a ScanNet++ probe scene. Thus the CIs likely overstate the precision of the headline gains if the selection procedure is itself part of the evaluation. This is a corollary of the first major comment, but it deserves separate emphasis because the CIs are used as the main evidence that the gains are 'not artifacts of a few scenes'. The CIs need to be either recomputed under a clean held-out calibration protocol or accompanied by a statement that they are conditional on a fixed, already-chosen τ.
minor comments (5)
  1. [§5, Baselines] The 'naive sparse-over-dense' baseline is described in the text as 'any lifted novel mask overwrites the dense label', but Table 3(a) and the Implementation section define it operationally as τ=0.2. A threshold of 0.2 still applies consistency filtering, so the label 'unconditional overwrite' is inaccurate. Please clarify the exact procedure used for this baseline.
  2. [§1 and §8] The introduction says the method uses a 'parameter-free rule' and the conclusion repeats the 'single threshold' claim, while the body correctly says the reconciliation has a single hyper-parameter τ. Please align the wording to avoid the impression that there is no threshold at all.
  3. [§8, Conclusion] The sentence 'bringing the harmonic mean to within 0.7 of trained GFS-VL' refers to the ScanNet++ result, but it appears in a paragraph that begins with ScanNet200. Please make the benchmark explicit to avoid ambiguity.
  4. [§5, Implementation] The SAM3 pipeline description says 'we keep, per frame and per concept, the highest-scoring returned mask at SAM3 confidence threshold 0.3'. It would help to state explicitly whether 0.3 is a threshold on the mask's confidence score, a minimum number of masks, or something else, and what happens to frames where no mask exceeds this value.
  5. [General] Several important details are deferred to 'the supplementary material' (the τ selection protocol, the leakage discussion, per-class breakdown Table S1, frame-rate/runtime breakdowns). Since the submitted version does not include a supplementary file, these details are currently missing. Please include the supplementary with the revision or move the essential parts into the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No equation-level or self-citation circularity; the τ calibration concern is a validation-selection risk, not a by-construction reduction.

full rationale

The paper's method is an empirical pipeline rather than a mathematical derivation, and its core reconciliation rule (Sec. 4) is a deterministic function of two frozen external models and one scalar threshold: y_i = n_hat if s_i >= tau else d_i. Neither s_i nor d_i is defined in terms of the reported mIoU; s_i is computed purely from per-frame SAM3 masks and view-visibility counts, and the dense prior is the frozen external RegionPLC model. The support-injection ablations use pre-committed null criteria, a reported development set, paired-bootstrap CIs, and disclose both the shared-feature limitation and the uninstantiated L3 locus. There is no self-citation chain: GFS-VL, RegionPLC, and SAM3 are all external prior work, and no 'uniqueness theorem' from the present authors is invoked. The only substantive concern is that the single hyper-parameter tau is calibrated per benchmark using evaluation-side information (the ScanNet++ probe scene and the ScanNet200 full-validation sweep, Table 3b), with the leakage defense deferred to a supplementary that is not present in the manuscript. That is a selection-bias / omitted-evidence problem, not circularity: tau is not fitted from or algebraically equivalent to the target mIoU, and the authors explicitly choose a value below the sweep optimum. The central comparison (dense prior vs. consensus-fused pipeline) remains an independent empirical measurement conditional on that disclosed operating point.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new entities, but relies on a small number of hand-picked parameters (primarily τ) and on the correctness of two frozen foundation models and the input sensor data. The threshold calibration is the most significant free parameter.

free parameters (3)
  • cross-view consistency threshold τ = 0.5 (ScanNet200), 0.15 (ScanNet++)
    Chosen per benchmark; on ScanNet++ fixed on a single probe scene; on ScanNet200 selected below the sweep optimum on the full validation set. The reported gains depend on this operating point.
  • SAM3 concept confidence threshold = 0.3
    Used to select masks per frame; no sensitivity analysis given.
  • frame subsampling stride = 20 (approx. 120 frames/scene)
    Subsampling of posed frames; ablation shows graceful degradation, but stride is a hand-picked choice.
assumptions (4)
  • domain assumption RegionPLC provides a dense but noisy 3D open-vocabulary prior over the evaluated classes.
    The method inherits RegionPLC's zero-shot capabilities; if RegionPLC is weak on novel classes, the dense branch is weak.
  • domain assumption SAM3 concept segmentation can reliably localize novel classes in 2D when prompted with class names.
    The sparse branch relies on SAM3's ability to map text to object masks; failure on stuff classes is acknowledged.
  • domain assumption The posed RGB frames and camera poses are aligned with the point cloud, and pinhole projection is sufficient.
    All sparse lifting depends on accurate poses and intrinsics; on ScanNet++ poses come from nerfstudio transforms.
  • domain assumption The GFS-PCS evaluation protocol and base/novel splits are as defined by GFS-VL.
    The paper adopts GFS-VL's benchmark without re-evaluating the protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Training-Free Open-Vocabulary 3D Point-Cloud Segmentation on the Generalized Few-Shot Benchmark." pith.science (2026). https://pith.science/paper/UBCLX2KZ

@misc{pith2026260715331,
  author       = {Pith},
  title        = {Pith review of: Training-Free Open-Vocabulary 3D Point-Cloud Segmentation on the Generalized Few-Shot Benchmark},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UBCLX2KZ}},
  note         = {Machine review of arXiv:2607.15331}
}
read the original abstract

Generalized few-shot 3D point-cloud segmentation (GFS-PCS) asks a model to segment a scene into many base classes seen at training time and a set of novel classes. The state of the art reaches novel classes by reconciling a dense but noisy 3D vision-language prior with the few-shot support, but it pays for this with base 3D labels, per-episode training, and the support annotations themselves. We ask how far the same reconciliation can go with none of these: no training, no 3D labels, and not even the few-shot support. We pair a frozen 3D vision-language model (RegionPLC) as a dense prior with a frozen promptable concept segmenter (SAM3), prompted by the bare novel class names and lifted from posed RGB views, and reconcile the two by cross-view consistency: a point becomes novel only when enough of the views that see it agree. On the ScanNet200 GFS-PCS benchmark this fully training-free, open-vocabulary pipeline improves novel mIoU by +2.6 over the training-free dense prior while holding base accuracy within 0.5, and recovers a third (33%) of the novel-class gap to the trained state of the art that uses far more supervision. We further show that injecting the few-shot support into the pipeline, as a fusion gate and as a prototypical dense classifier, adds nothing over consistency alone and in fact degrades it through the classifier, which is why the method needs no support at all. On the harder ScanNet++ benchmark, where the dense prior is far weaker on novel classes, the same pipeline nearly doubles novel mIoU (+15.7, from 16.2 to 31.9) at a 1.7 base cost, lifting the harmonic mean from 21.5 to 31.1

Figures

Figures reproduced from arXiv: 2607.15331 by the authors.

Figure 1
Figure 1. Prior GFS-PCS vs. ours. Top: the state of the art (GFS-VL) trains a 3D segmentor on base 3D labels and fine-tunes a registration step on the few-shot support. Bottom: we pair a frozen 3D vision-language model (RegionPLC) with a frozen promptable concept segmenter (SAM3-PCS) prompted by the bare novel class names, and reconcile their predictions by cross-view consistency. We use no training, no 3D labels, and not eve… view at source ↗
Figure 2
Figure 2. Pipeline overview (all components frozen). [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results on ScanNet++ (novel classes colored, base and unla￾beled points grey; near-top-down view with the ceiling removed for visibility). Columns: input RGB, ground-truth novel labels, the dense prior (RegionPLC), and Ours (cross￾view consensus, τ=0.15). Consensus recovers small text-promptable novel objects the dense prior misses—the windowsill and window frame (row 2) and the blinds strip (row 3)—at t… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 4 linked inside Pith

  1. [1]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 5

    Alayrac, J.B., Donahue, J., Luc, P., Miech, A., Barr, I., Hasson, Y., Lenc, K., Mensch, A., Millican, K., Reynolds, M., et al.: Flamingo: a visual language model Training-Free Open-Vocabulary GFS-PCS 15 for few-shot learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 5

    2022

  2. [2]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 2, 3, 5, 7, 8, 9, 11

    An, Z., Sun, G., Liu, Y., Li, R., Han, J., Konukoglu, E., Belongie, S.: Generalized few-shot 3d point cloud segmentation with vision-language model. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 2, 3, 5, 7, 8, 9, 11

    2025

  3. [3]

    In: Inter- national Conference on Learning Representations (ICLR) (2025) 3

    An, Z., Sun, G., Liu, Y., Li, R., Wu, M., Cheng, M.M., Konukoglu, E., Belongie, S.: Multimodality helps few-shot 3d point cloud semantic segmentation. In: Inter- national Conference on Learning Representations (ICLR) (2025) 3

    2025

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3

    An, Z., Sun, G., Liu, Y., Liu, F., Wu, Z., Wang, D., Van Gool, L., Belongie, S.: Rethinking few-shot 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3

    2024

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 5

    Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I., Fischer, M., Savarese, S.: 3d semantic parsing of large-scale indoor spaces. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016) 5

    2016

  6. [6]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 5

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2020) 5

    2020

  7. [7]

    org/abs/2511.167192, 5, 7

    Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Alwala, K.V., Khedr, H., Huang, A., Lei, J., Ma, T., Guo, B., Kalla, A., Marks, M., Greer, J., Wang, M., Sun, P., Rädle, R., Afouras, T., Mavroudi, E., Xu, K., Wu, T.H., Zhou, Y., Momeni, L., Hazra, R., Ding, S., Vaze, S., Porcher, F., Li, F., Li, S., Kamath, A., Cheng, H.K., ...

    arXiv 2025

  8. [8]

    In: British Machine Vision Conference (BMVC) (2021) 3

    Cermelli, F., Mancini, M., Xian, Y., Akata, Z., Caputo, B.: Prototype-based in- cremental few-shot semantic segmentation. In: British Machine Vision Conference (BMVC) (2021) 3

    2021

Show all 51 references
  1. [9]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3

    Chen, R., Liu, Y., Kong, L., Zhu, X., Ma, Y., Li, Y., Hou, Y., Qiao, Y., Wang, W.: CLIP2Scene: Towards label-efficient 3d scene understanding by CLIP. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3

    2023

  2. [10]

    In: Proceedings of the European Con- ference on Computer Vision (ECCV) (2020) 5

    Chen, Y., Hu, V.T., Gavves, E., Mensink, T., Mettes, P., Yang, P., Snoek, C.G.: PointMixup: Augmentation for point clouds. In: Proceedings of the European Con- ference on Computer Vision (ECCV) (2020) 5

    2020

  3. [11]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 5

    Choy,C.,Gwak,J.,Savarese,S.:4dspatio-temporalConvNets:Minkowskiconvolu- tional neural networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 5

    2019

  4. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 5

    Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scan- Net: Richly-annotated 3d reconstructions of indoor scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2017) 5

    2017

  5. [13]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2023) 3

    Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: PLA: Language-driven open-vocabulary 3d scene understanding. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2023) 3

    2023

  6. [14]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2024) 3 16 S.K.Gah and E.Owusu

    Ding, R., Yang, J., Xue, C., Zhang, W., Bai, S., Qi, X.: Lowis3D: Language-driven open-world instance-level 3d scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2024) 3 16 S.K.Gah and E.Owusu

    2024

  7. [15]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 5

    Ghiasi,G.,Gu,X.,Cui,Y.,Lin,T.Y.:Scalingopen-vocabularyimagesegmentation with image-level labels. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 5

    2022

  8. [16]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2018) 5

    Graham, B., Engelcke, M., van der Maaten, L.: 3d semantic segmentation with sub- manifold sparse convolutional networks. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2018) 5

    2018

  9. [17]

    In: International Joint Conference on Artificial Intelligence (IJCAI) (2024) 5

    He,Q.,Peng,J.,Jiang,Z.,Wu,K.,Ji,X.,Zhang,J.,Wang,Y.,Wang,C.,Chen,M., Wu, Y.: Unim-ov3d: Uni-modality open-vocabulary 3d scene understanding with fine-grained feature representation. In: International Joint Conference on Artificial Intelligence (IJCAI) (2024) 5

    2024

  10. [18]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 3

    Huang, T., Dong, B., Yang, Y., Huang, X., Lau, R.W., Ouyang, W., Zuo, W.: CLIP2Point: Transfer CLIP to point cloud classification with image-depth pre- training. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 3

    2023

  11. [19]

    In: European Conference on Computer Vision (ECCV) (2024) 4

    Huang, Z., Wu, X., Chen, X., Zhao, H., Zhu, L., Lasenby, J.: Openins3d: Snap and lookup for 3d open-vocabulary instance segmentation. In: European Conference on Computer Vision (ECCV) (2024) 4

    2024

  12. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3

    Jiang, L., Shi, S., Schiele, B.: Open-vocabulary 3d semantic segmentation with foundation models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3

    2024

  13. [21]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 5

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollár, P., Girshick, R.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 5

    2023

  14. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 5

    Lai, X., Liu, J., Jiang, L., Wang, L., Zhao, H., Liu, S., Qi, X., Jia, J.: Stratified transformer for 3d point cloud segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2022) 5

    2022

  15. [23]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 5

    Lee, J., Park, C., Choe, J., Wang, Y.C.F., Kautz, J., Cho, M., Choy, C.: Mo- saic3d: Foundation dataset and model for open-vocabulary 3d segmentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2025) 5

    2025

  16. [24]

    Lee,S.,Zhao,Y.,Lee,G.H.:Segmentany3dobjectwithlanguage.In:International Conference on Learning Representations (ICLR) (2025) 5

    2025

  17. [25]

    In: International Conference on Learning Representations (ICLR) (2022) 5

    Li, B., Weinberger, K.Q., Belongie, S., Koltun, V., Ranftl, R.: Language-driven semantic segmentation. In: International Conference on Learning Representations (ICLR) (2022) 5

    2022

  18. [26]

    In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition

    Li, L.H., Zhang, P., Zhang, H., Yang, J., Li, C., Zhong, Y., Wang, L., Yuan, L., Zhang, L., Hwang, J.N., et al.: Grounded language-image pre-training. In: Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10965–10975 (2022) 5

    2022

  19. [27]

    In: European Conference on Computer Vision

    Li, R., Zhang, Z., He, C., Ma, Z., Patel, V.M., Zhang, L.: Dense multimodal align- ment for open-vocabulary 3d scene understanding. In: European Conference on Computer Vision. pp. 416–434. Springer (2024) 3, 5

    2024

  20. [28]

    In: International Conference on 3D Vision (3DV) (2021) 5

    Nekrasov, A., Schult, J., Litany, O., Leibe, B., Engelmann, F.: Mix3D: Out-of- context data augmentation for 3d scenes. In: International Conference on 3D Vision (3DV) (2021) 5

    2021

  21. [29]

    In: Proceedings of the ACM International Conference on Multimedia (ACM MM) (2023) 3 Training-Free Open-Vocabulary GFS-PCS 17

    Ning, Z., Tian, Z., Lu, G., Pei, W.: Boosting few-shot 3d point cloud segmen- tation via query-guided enhancement. In: Proceedings of the ACM International Conference on Multimedia (ACM MM) (2023) 3 Training-Free Open-Vocabulary GFS-PCS 17

    2023

  22. [30]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3

    Peng, S., Genova, K., Jiang, C.M., Tagliasacchi, A., Pollefeys, M., Funkhouser, T.: OpenScene: 3d scene understanding with open vocabularies. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2023) 3

    2023

  23. [31]

    arXiv preprint arXiv:2306.14824 (2023) 5

    Peng, Z., Wang, W., Dong, L., Hao, Y., Huang, S., Ma, S., Wei, F.: Kosmos- 2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824 (2023) 5

    2023 arXiv

  24. [32]

    In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 3, 5

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning trans- ferable visual models from natural language supervision. In: Proceedings of the International Conference on Machin...

    2021

  25. [33]

    arXiv preprint arXiv:2408.00714 (2024) 5

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., et al.: SAM 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024) 5

    2024 arXiv

  26. [34]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 3, 5

    Rozenberszki, D., Litany, O., Dai, A.: Language-grounded indoor 3d semantic seg- mentation in the wild. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 3, 5

    2022

  27. [35]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2017) 3, 9, 11

    Snell, J., Swersky, K., Zemel, R.S.: Prototypical networks for few-shot learning. In: Advances in Neural Information Processing Systems (NeurIPS) (2017) 3, 9, 11

    2017

  28. [36]

    arXiv preprint arXiv:2405.15580 (2024) 4

    Tai, H., He, Q., Zhang, J., Qian, Y., Zhang, Z., Hu, X., Li, X., Wang, Y., Liu, Y.: Open-vocabulary sam3d: Towards training-free open-vocabulary 3d scene under- standing. arXiv preprint arXiv:2405.15580 (2024) 4

    2024 arXiv

  29. [37]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 3

    Takmaz, A., Fedele, E., Sumner, R.W., Pollefeys, M., Tombari, F., Engelmann, F.: OpenMask3D: Open-vocabulary 3d instance segmentation. In: Advances in Neural Information Processing Systems (NeurIPS) (2023) 3

    2023

  30. [38]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019) 5

    Thomas, H., Qi, C.R., Deschaud, J.E., Marcotegui, B., Goulette, F., Guibas, L.J.: KPConv: Flexible and deformable convolution for point clouds. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2019) 5

    2019

  31. [39]

    In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) (2022) 3

    Tian, Z., Lai, X., Jiang, L., Liu, S., Shu, M., Zhao, H., Jia, J.: Generalized few-shot semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR) (2022) 3

    2022

  32. [40]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2024) 3

    Tsai, C.J., Chen, H.T., Liu, T.L.: Pseudo-embedding for generalized few-shot 3d segmentation. In: Proceedings of the European Conference on Computer Vision (ECCV) (2024) 3

    2024

  33. [41]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 5

    Wu, W., Qi, Z., Fuxin, L.: PointConv: Deep convolutional networks on 3d point clouds. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2019) 5

    2019

  34. [42]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3, 5

    Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point transformer V3: Simpler, faster, stronger. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3, 5

    2024

  35. [43]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 3

    Xu, Y., Hu, C., Zhao, N., Lee, G.H.: Generalized few-shot point cloud segmentation via geometric words. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 3

    2023

  36. [44]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 2, 3, 7

    Yang, J., Ding, R., Deng, W., Wang, Z., Qi, X.: RegionPLC: Regional point- language contrastive learning for open-world 3d scene understanding. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 2, 3, 7

    2024

  37. [45]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 5

    Yao, L., Han, J., Wen, Y., Liang, X., Xu, D., Zhang, W., Li, Z., Xu, C., Xu, H.: DetCLIP: Dictionary-enriched visual-concept paralleled pre-training for open- 18 S.K.Gah and E.Owusu world detection. In: Advances in Neural Information Processing Systems (NeurIPS) (2022) 5

    2022

  38. [46]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 3, 5

    Yeshwanth,C.,Liu,Y.C.,Nießner,M.,Dai,A.:ScanNet++:Ahigh-fidelitydataset of 3d indoor scenes. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (2023) 3, 5

    2023

  39. [47]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (2023) 3

    Zhang, J., Dong, R., Ma, K.: CLIP-FO3D: Learning free open-world 3d scene rep- resentations from 2d dense CLIP. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops (2023) 3

    2023

  40. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 3

    Zhao, N., Chua, T.S., Lee, G.H.: Few-shot 3d point cloud semantic segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2021) 3

    2021

  41. [49]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 5

    Zhou, C., Loy, C.C., Dai, B.: Extract free dense labels from CLIP. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 5

    2022

  42. [50]

    In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 5

    Zhou, X., Girdhar, R., Joulin, A., Krähenbühl, P., Misra, I.: Detecting twenty- thousand classes using image-level supervision. In: Proceedings of the European Conference on Computer Vision (ECCV) (2022) 5

    2022

  43. [51]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3

    Zhu, X., Zhang, R., He, B., Guo, Z., Liu, J., Xiao, H., Fu, C., Dong, H., Gao, P.: No time to train: Empowering non-parametric networks for few-shot 3d scene segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024) 3

    2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.