Pith. sign in

REVIEW 3 major objections 5 minor 28 references

CDIS claims that closing the loop between 2D mask tracking and 3D superpoint anchoring yields globally consistent, class-agnostic 3D instance segmentation without any 3D-specific training.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:03 UTC pith:PVCVP5RY

load-bearing objection Useful but uneven zero-shot 3D instance segmentation paper: the tracking-plus-superpoint loop is a genuine innovation, yet the superpoint construction is underspecified and the evaluation lacks ablations and error bars. the 3 major comments →

arxiv 2607.17778 v1 pith:PVCVP5RY submitted 2026-07-20 cs.CV cs.AI

CDIS: Cross-Dimensional Class-Agnostic 3D Instance Segmentation via 2D Mask Tracking and 3D-2D Projection Merging

Juno Kim , Hye-Jung Yoon , Yesol Park , Byoung-Tak Zhang This is my paper
classification cs.CV cs.AI
keywords zero-shot 3D instance segmentationclass-agnostic segmentation2D mask tracking3D superpointscross-dimensional fusionRGB-D perceptiontemporal consistencyprojection merging
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to show that the main weakness of zero-shot 3D instance segmentation — fragmented object identities caused by merging per-frame 2D masks independently — can be fixed by explicitly tracking instances in 2D over time and using 3D superpoints as spatial anchors. It proposes CDIS, which runs 2D instance tracking with depth warping and IoU matching, then associates tracked masks with precomputed 3D superpoints, merges instances across frames via 3D overlap and temporal co-occurrence, and iterates. The claim is that this cross-dimensional feedback loop yields globally consistent 3D instance labels without any 3D-specific training. If true, a robot could segment novel objects from posed RGB-D streams without labeled 3D data, with improved temporal consistency compared to existing zero-shot methods.

Core claim

The central claim is that a closed loop between 2D and 3D produces better zero-shot 3D instance segmentation than one-way projection: 2D tracking over multiple past frames corrects transient segmentation errors, and 3D superpoints resolve spatial ambiguities; the two domains reinforce each other. On the larger of two indoor benchmark validation sets, CDIS with one pretrained 2D mask generator reports 33.2 average precision versus 31.5 for the best prior zero-shot method, and on the second benchmark it reports 28.2 versus 27.9, while requiring no 3D training.

What carries the argument

The load-bearing object is the set of 3D superpoints — geometrically coherent clusters of the scene point cloud — used as spatial anchors. Each superpoint is projected onto every frame, assigned the 2D instance label with greatest footprint overlap, and then used to compute instance similarity across frames through 3D IoU over superpoint membership. This converts 2D tracking into a 3D-consistent voting procedure, and duplicate superpoint assignments are resolved by 3D overlap and temporal co-occurrence analysis.

Load-bearing premise

The whole pipeline rests on the assumption that the 3D superpoints are geometrically coherent and align with the 2D instance boundaries; if a superpoint straddles multiple objects or projects ambiguously (due to depth noise, occlusions, reflections), errors propagate to all merging steps.

What would settle it

A concrete test would be to construct scenes with known ground-truth instances where the superpoint boundaries deliberately straddle object boundaries (e.g., two objects touching or one object with a reflective surface), and check whether CDIS's AP drops below simpler per-frame projection methods; if the superpoint association is the bottleneck, errors would cluster at those boundaries.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Zero-shot 3D instance segmentation can be improved by exploiting temporal consistency rather than treating frames independently.
  • The cross-dimensional loop reduces fragmentation and identity switches in 3D instances, which matters for robotic manipulation and navigation.
  • Performance scales across different 2D mask generators: with one generator CDIS reaches 33.2 AP on the first benchmark, and 30.1 with another, showing the framework is not tied to a specific 2D segmenter.
  • Without any 3D supervision, CDIS approaches the accuracy of a fully supervised 3D model (33.2 AP versus 39.7 AP) on the same benchmark, narrowing the gap between zero-shot and trained methods.
  • Real-world RGB-D scenes reconstructed with visual SLAM and signed-distance-function integration can be segmented class-agnostically, with reduced accuracy only near strongly reflective surfaces.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the tracking loop proves robust, the same cross-dimensional feedback could be applied to video object segmentation, where 3D anchors constrain drifting 2D masks.
  • The method's overall ceiling appears directly tied to superpoint quality; better geometric grouping should translate immediately into higher AP without changing the 2D stage.
  • Because the framework is agnostic to the 2D mask predictor, future stronger 2D segmenters should automatically lift the 3D results, making the approach a forward-compatible wrapper.
  • The temporal co-occurrence analysis might be extended to actively split instances that reappear after occlusion, a scenario not isolated in the current benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes CDIS, a zero-shot, class-agnostic 3D instance segmentation method that combines 2D instance masks from SAM or CropFormer with 3D superpoints. It first tracks 2D masks across posed RGB-D frames via depth warping and IoU matching, then associates each superpoint with the dominant 2D instance over frames, merges instances using 3D IoU, and consolidates duplicates using an overlap and temporal co-occurrence criterion. The final output assigns each superpoint an instance ID by voting. Evaluation is reported on ScanNet200 and ScanNet++ validation sets, with the best configuration (CropFormer-based) reaching 33.2 AP on ScanNet200 and 28.2 AP on ScanNet++, together with qualitative real-world results.

Significance. The cross-dimensional feedback idea is timely and plausible: using temporally tracked 2D masks to label 3D superpoints could reduce fragmentation and drift in projection-based zero-shot pipelines. The framework is training-free and leverages strong pretrained 2D segmenters, and the evaluation uses standard public benchmarks. If the reported results are reproducible and robust to the many hand-set thresholds, the method would be a useful contribution to open-world/organized 3D perception. However, the empirical support is currently fragile: the crucial 3D superpoint construction is not specified, no ablation or sensitivity analysis is provided, and the reported gains are often small or statistically unsupported. The paper does not ship code, so independent verification is not possible at this stage. The central idea is worth pursuing, but the manuscript needs substantial revision to make the claims defensible.

major comments (3)
  1. [Section III-B, 3D Superpoint Projection] The paper says superpoints are 'obtained from [26]' (Felzenszwalb-Huttenlocher, an image segmentation method) and 'adapted to the scene point cloud,' but gives no details: what graph is built (k-NN? radius?), what edge weights are used (position? color? normals?), what scale parameter or minimum component size is chosen, and how the original 2D graph-based algorithm is generalized to 3D. This is load-bearing because the final output is a superpoint-level labeling assigned by voting in Sec. III-C. If superpoints straddle object boundaries, Eq. (5)'s argmax assignment commits a boundary-spanning superpoint to one object and the error propagates through Eqs. (6)-(8). The paper provides no superpoint-quality analysis, no scale sensitivity study, and no robustness test with noisy depth, so the reported 33.2 / 28.2 AP cannot be independently verified or trusted.
  2. [Section IV-A, Implementation Details; Tables I and II] Five free thresholds are hand-set (q_max=5, tau_IoU2D=0.8, tau_IoU3D=0.6, tau_IoMin3D=0.8, tau_co=0.5) with no ablation or sensitivity analysis. Since these thresholds directly determine when tracks are merged and when duplicates are resolved, the reported numbers are of unknown stability. Furthermore, all results are single runs on validation sets with no error bars or multiple seeds. The claimed advantage over MaskClustering on ScanNet++ is only 0.3 AP (28.2 vs 27.9) and AP25 is actually lower (54.3 vs 54.7); such a marginal difference cannot be asserted as a 'higher accuracy and consistency' result without variance estimates or significance testing.
  3. [Section IV-B, Quantitative Results] The comparison protocol is under-specified. It is unclear whether the 2D mask generators (SAM/CropFormer) use the same settings, whether the same superpoints are used across methods, and whether OV-MAP [14] is the exact version from the cited IROS paper or a re-implementation. The paper reports only validation-set numbers, no test-set or cross-scene breakdown. Given that the two strongest baselines (Open3DIS, MaskClustering) are state of the art, the paper should provide a careful Apple-to-apples comparison and a per-scene analysis, especially for ScanNet++ where the margin is tiny.
minor comments (5)
  1. [Eqs. (1)-(3)] The warping equation uses homogeneous coordinates but 'P_tp' and 'P_tp->t' are not explicitly declared as homogeneous; Eq. (2) also has potentially ambiguous transposes. Eq. (3) does not specify how pixels with no warped point (holes, occlusion) are handled. Clarify these details.
  2. [Section III-C, Temporal Co-Occurrence] The rule 'If CO(u,v)>tau_co, the two instances are likely distinct objects' is counterintuitive and could be stated more clearly: the reader may expect high co-occurrence to indicate the same object. The intuitive interpretation of the threshold should be given.
  3. [Table II] The MaskClustering row is formatted as '42.854.7' in the text. This is likely a missing separator between AP50 and AP25. Please correct the typesetting.
  4. [Section IV-D, Real-World Experiments] The real-world evaluation is purely qualitative. Consider reporting at least IoU-based counts or a qualitative error analysis. Also, the statement that segmentation accuracy decreased near reflective windows is not quantified; this could be a useful place to diagnose where the tracking/projection pipeline fails.
  5. [Introduction / Fig. 1] The caption and the introduction mention a 'feedback loop' between 2D and 3D, but the method as described is a one-way pipeline from 2D tracking to 3D consolidation. If there is an actual feedback mechanism (e.g., 3D-consolidated labels are fed back to correct 2D masks), it is not reflected in the equations in Section III. Please de-emphasize 'feedback' or make the loop explicit.

Circularity Check

0 steps flagged

No significant circularity: CDIS is an empirical pipeline whose claims are evaluated on external benchmarks.

full rationale

CDIS is an empirical pipeline, not a derivation. The final 3D superpoint labels are computed from 2D instance masks via projection association (Eq. 5), pairwise merging (Eq. 6), and voting; this is algorithmic construction, not a prediction derived from the output. The reported AP numbers are evaluated on external benchmark ground truth (ScanNet200, ScanNet++), so the central claim is externally falsifiable. The only self-citations are [14] (OV-MAP) and [23] (DA-Fusion); they appear as related-work/comparison references, and no load-bearing premise is justified by them. The underspecified 'adapted to the scene point cloud' superpoint step ([26]) is a reproducibility/soundness concern, not a circularity: the paper does not define superpoints in terms of the target labels, nor fit a parameter to the benchmark. No equation in the paper reduces to its own input by construction, and no fitted parameter is renamed as a prediction. The central claim is therefore self-contained against the benchmark evaluations, and no circular step is present.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on five hand-set thresholds (no ablations) and four domain assumptions about 2D segmentation stability, 3D superpoint coherence, and the temporal co-occurrence logic. No new physical or theoretical entities are proposed.

free parameters (5)
  • frame queue length q_max = 5
    Limits how many previous frames are considered for 2D tracking; set by hand in Section IV-A with no sensitivity analysis.
  • 2D IoU threshold tau_IoU2D = 0.8
    Threshold for matching instance labels across warped frames; hand-set in Section IV-A.
  • 3D IoU threshold tau_IoU3D = 0.6
    Threshold for merging instance labels based on shared superpoints between adjacent frames; hand-set in Section IV-A.
  • 3D IoMin threshold tau_IoMin3D = 0.8
    Threshold for detecting overlapping instances during consolidation; hand-set in Section IV-A.
  • temporal co-occurrence threshold tau_co = 0.5
    Threshold distinguishing distinct objects that co-occur from the same object seen at different times; hand-set in Section IV-A.
axioms (4)
  • standard math Perspective camera model with known intrinsic matrix K and accurate poses C_t
    Used in Eqs. (1)-(3) for warping and projection; standard for RGB-D processing but assumes calibration accuracy.
  • domain assumption Pretrained 2D segmenters (SAM / CropFormer) yield instance-stable masks across frames of the same object
    The whole tracking stage depends on the 2D model giving class-agnostic masks whose identity persists, with only partial correction from multi-frame matching (Fig. 3).
  • ad hoc to paper Felzenszwalb-Huttenlocher [26] produces instance-aligned 3D superpoints
    Section III-B invokes a 2D image over-segmentation method for 3D superpoints without specifying the adaptation or verifying instance alignment; this is a load-bearing ingredient for Eq. (5).
  • ad hoc to paper Temporal co-occurrence heuristic distinguishes distinct objects from same-object observations
    Eq. (8): CO(u,v)>tau_co is interpreted as 'distinct objects whose overlap is noise', but a single object that is intermittently occluded can violate this, and two adjacent objects observed together will be kept separate.

pith-pipeline@v1.3.0-alltime-deepseek · 8245 in / 14240 out tokens · 140434 ms · 2026-08-01T17:03:13.350979+00:00 · methodology

0 comments
read the original abstract

Class-agnostic 3D instance segmentation is critical for robotic systems operating in unknown environments, enabling perception of previously unseen objects for reliable manipulation and navigation. Existing approaches typically project per-frame 2D instance masks into 3D and merge them, which often breaks object identities across time and yields fragmented 3D instances. We introduce Cross-Dimensional Class-Agnostic 3D Instance Segmentation (CDIS), a zero-shot framework that explicitly tracks 2D instance masks across frames and associates them with 3D superpoints, creating a feedback loop between 2D and 3D. This cross-dimensional reasoning links temporally stable 2D tracks with spatially coherent 3D regions, producing globally consistent 3D instance labels without any 3D-specific training. Experiments on benchmark datasets demonstrate that CDIS achieves higher accuracy and consistency than state-of-the-art zero-shot methods, while remaining efficient and scalable to diverse real-world environments.

Figures

Figures reproduced from arXiv: 2607.17778 by Byoung-Tak Zhang, Hye-Jung Yoon, Juno Kim, Yesol Park.

Figure 1
Figure 1. Figure 1: Overview of CDIS. Top image illustrates a reconstructed real￾world 3D scene generated from sequentially captured RGB-D frames at timestamps t1,t2,t3. The bottom images show the progression of 2D instance segmentation results over time, demonstrating our method, which improves segmentation quality across frames through mask tracking. CDIS integrates segmentation cues from both 2D and 3D spaces to enhance cl… view at source ↗
Figure 2
Figure 2. Figure 2: Overall Framework of CDIS. The proposed pipeline for zero-shot, class-agnostic 3D instance segmentation. Posed RGB-D frames are input into a 2D mask predictor to generate instance masks for each frame, which are tracked over time using depth-based projection, frame warping, and 2D IoU matching. Tracked masks are then associated with pre-computed 3D superpoints, enabling spatio-temporal merging of instances… view at source ↗
Figure 3
Figure 3. Figure 3: Handling 2D Segmentation Failures with Multi-Frame Tracking. When generating the mask for frame t, using only the warped mask from frame t −1 propagates an error where the chair and table are merged (red box). CDIS prevents this by also referencing frame t−2, where the objects are correctly separated, and restores proper instance IDs in frame t. By leveraging multiple past frames rather than only the adjac… view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of Class-Agnostic 3D Instance Segmentation on ScanNet200. Visualization of 3D instance segmentation results. From left to right: input, ground truth (GT), Mask3D [5], SAM3D [12], and our method (CDIS). CDIS demonstrates improved segmentation accuracy, particularly in distinguishing object boundaries and reducing over-segmentation, as highlighted with red circles. C. Qualitative Results As shown … view at source ↗
Figure 5
Figure 5. Figure 5: Example of Real-World Class-Agnostic 3D Instance Segmen￾tation. CDIS performance on real-world data, demonstrating class-agnostic instance segmentation across large and diverse indoor environments, includ￾ing an office (red), a bedroom (yellow), and a kitchen (blue). Our method effectively segments cluttered and complex scenes without 3D training. scene geometry was reconstructed using a Truncated Signed D… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 8 linked inside Pith

  1. [1]

    Language-grounded indoor 3d semantic segmentation in the wild,

    D. Rozenberszki, O. Litany, and A. Dai, “Language-grounded indoor 3d semantic segmentation in the wild,” inProceedings of the European Conference on Computer Vision (ECCV), 2022

    2022

  2. [2]

    Scannet++: A high- fidelity dataset of 3d indoor scenes,

    C. Yeshwanth, Y .-C. Liu, M. Nießner, and A. Dai, “Scannet++: A high- fidelity dataset of 3d indoor scenes,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 12–22, 2023

    2023

  3. [3]

    Matterport3d: Learning from rgb-d data in indoor environments,

    A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Niessner, M. Savva, S. Song, A. Zeng, and Y . Zhang, “Matterport3d: Learning from rgb-d data in indoor environments,”arXiv preprint arXiv:1709.06158, 2017

    Pith/arXiv arXiv 2017

  4. [4]

    The replica dataset: A digital replica of indoor spaces,

    J. Straub, T. Whelan, L. Ma, Y . Chen, E. Wijmans, S. Green, J. J. Engel, R. Mur-Artal, C. Ren, S. Verma,et al., “The replica dataset: A digital replica of indoor spaces,”arXiv preprint arXiv:1906.05797, 2019

    Pith/arXiv arXiv 1906

  5. [5]

    Mask3d: Mask transformer for 3d semantic instance segmentation,

    J. Schult, F. Engelmann, A. Hermans, O. Litany, S. Tang, and B. Leibe, “Mask3d: Mask transformer for 3d semantic instance segmentation,” in2023 IEEE International Conference on Robotics and Automation (ICRA), pp. 8216–8223, IEEE, 2023

    2023

  6. [6]

    Openmask3d: Open-vocabulary 3d instance segmen- tation,

    A. Takmaz, E. Fedele, R. W. Sumner, M. Pollefeys, F. Tombari, and F. Engelmann, “Openmask3d: Open-vocabulary 3d instance segmen- tation,”arXiv preprint arXiv:2306.13631, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    Oneformer3d: One transformer for unified point cloud segmentation,

    M. Kolodiazhnyi, A. V orontsova, A. Konushin, and D. Rukhovich, “Oneformer3d: One transformer for unified point cloud segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20943–20953, 2024

    2024

  8. [8]

    Ovir-3d: Open-vocabulary 3d instance retrieval without training on 3d data,

    S. Lu, H. Chang, E. P. Jing, A. Boularias, and K. Bekris, “Ovir-3d: Open-vocabulary 3d instance retrieval without training on 3d data,” in Conference on Robot Learning, pp. 1610–1620, PMLR, 2023

    2023

  9. [9]

    Unscene3d: Unsupervised 3d instance segmentation for indoor scenes,

    D. Rozenberszki, O. Litany, and A. Dai, “Unscene3d: Unsupervised 3d instance segmentation for indoor scenes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 19957–19967, 2024

    2024

  10. [10]

    Sampro3d: Locating sam prompts in 3d for zero-shot scene segmentation,

    M. Xu, X. Yin, L. Qiu, Y . Liu, X. Tong, and X. Han, “Sampro3d: Locating sam prompts in 3d for zero-shot scene segmentation,”arXiv preprint arXiv:2311.17707, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d in- stance segmentation,

    M. Yan, J. Zhang, Y . Zhu, and H. Wang, “Maskclustering: View consensus based mask graph clustering for open-vocabulary 3d in- stance segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 28274–28284, 2024

    2024

  12. [12]

    Sam3d: Segment anything in 3d scenes,

    Y . Yang, X. Wu, T. He, H. Zhao, and X. Liu, “Sam3d: Segment anything in 3d scenes,”arXiv preprint arXiv:2306.03908, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Sam-guided graph cut for 3d instance segmentation,

    H. Guo, H. Zhu, S. Peng, Y . Wang, Y . Shen, R. Hu, and X. Zhou, “Sam-guided graph cut for 3d instance segmentation,”arXiv preprint arXiv:2312.08372, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Ov-map: Open- vocabulary zero-shot 3d instance segmentation map for robots,

    J. Kim, Y . Park, H.-J. Yoon, and B.-T. Zhang, “Ov-map: Open- vocabulary zero-shot 3d instance segmentation map for robots,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 13780–13786, IEEE, 2024

    2024

  15. [15]

    Sai3d: Segment any instance in 3d scenes,

    Y . Yin, Y . Liu, Y . Xiao, D. Cohen-Or, J. Huang, and B. Chen, “Sai3d: Segment any instance in 3d scenes,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3292– 3302, 2024

    2024

  16. [16]

    Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance,

    P. Nguyen, T. D. Ngo, E. Kalogerakis, C. Gan, A. Tran, C. Pham, and K. Nguyen, “Open3dis: Open-vocabulary 3d instance segmentation with 2d mask guidance,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 4018–4028, 2024

    2024

  17. [17]

    Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels,

    R. Huang, S. Peng, A. Takmaz, F. Tombari, M. Pollefeys, S. Song, G. Huang, and F. Engelmann, “Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels,”arXiv preprint arXiv:2312.17232, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Openins3d: Snap and lookup for 3d open-vocabulary instance seg- mentation,

    Z. Huang, X. Wu, X. Chen, H. Zhao, L. Zhu, and J. Lasenby, “Openins3d: Snap and lookup for 3d open-vocabulary instance seg- mentation,” inEuropean Conference on Computer Vision, pp. 169– 185, Springer, 2024

    2024

  19. [19]

    Per-pixel classification is not all you need for semantic segmentation,

    B. Cheng, A. Schwing, and A. Kirillov, “Per-pixel classification is not all you need for semantic segmentation,”Advances in neural information processing systems, vol. 34, pp. 17864–17875, 2021

    2021

  20. [20]

    Masked-attention mask transformer for universal image segmenta- tion,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmenta- tion,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 1290–1299, 2022

    2022

  21. [21]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo,et al., “Segment anything,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 4015–4026, 2023

    2023

  22. [22]

    High-quality entity segmentation,

    L. Qi, J. Kuen, W. Guo, T. Shen, J. Gu, J. Jia, Z. Lin, and M.-H. Yang, “High-quality entity segmentation,”arXiv preprint arXiv:2211.05776, 2022

    Pith/arXiv arXiv 2022

  23. [23]

    DA-Fusion: De- formable attention-based rgb-d fusion transformer for unseen object instance segmentation,

    Y . Park, H.-J. Yoon, J. Kim, and B.-T. Zhang, “DA-Fusion: De- formable attention-based rgb-d fusion transformer for unseen object instance segmentation,” in2025 IEEE International Conference on Robotics and Automation (ICRA), pp. 7490–7496, IEEE, 2025

    2025

  24. [24]

    3d-sis: 3d semantic instance segmen- tation of rgb-d scans,

    J. Hou, A. Dai, and M. Nießner, “3d-sis: 3d semantic instance segmen- tation of rgb-d scans,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 4421–4430, 2019

    2019

  25. [25]

    Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution,

    T. D. Ngo, B.-S. Hua, and K. Nguyen, “Isbnet: a 3d point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13550– 13559, 2023

    2023

  26. [26]

    Efficient graph-based image segmentation,

    P. F. Felzenszwalb and D. P. Huttenlocher, “Efficient graph-based image segmentation,”International journal of computer vision, vol. 59, no. 2, pp. 167–181, 2004

    2004

  27. [27]

    Rtab-map as an open-source lidar and visual simultaneous localization and mapping library for large-scale and long-term online operation,

    M. Labb ´e and F. Michaud, “Rtab-map as an open-source lidar and visual simultaneous localization and mapping library for large-scale and long-term online operation,”Journal of field robotics, vol. 36, no. 2, pp. 416–446, 2019

    2019

  28. [28]

    Kinectfusion: Real-time dense surface mapping and tracking,

    R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohi, J. Shotton, S. Hodges, and A. Fitzgibbon, “Kinectfusion: Real-time dense surface mapping and tracking,” in 2011 10th IEEE international symposium on mixed and augmented reality, pp. 127–136, Ieee, 2011

    2011