Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

3D part segmentation can be solved as promptable multi-view 2D mask prediction: a click or box on any of 12 rendered views lifts SAM2 masks to consistent 3D parts, reaching 84.06 mIoU on PartObjaverse-Tiny and 74.42 on PartNetE.

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 ↗

A prompt-controllable 3D part segmentation method that adapts SAM2 with LoRA and geometry fusion on rendered normal and point maps, then back-projects multi-view masks to the mesh.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Prompt-controllable 3D part segmentation by lifting SAM2 is a good idea with a solid ablation story, but the oracle view-selection protocol makes the SOTA claim unsupported as written. the 3 major comments →

arxiv 2508.14036 v2 pith:6XDW4JQB submitted 2025-08-19 cs.CV cs.AI

GeoSAM2: Unleashing the Power of SAM2 for 3D Part Segmentation

Ken Deng , Yunhan Yang , Jingxiang Sun , Xihui Liu , Yebin Liu , Ding Liang , Yan-Pei Cao This is my paper
classification cs.CV cs.AI
keywords 3D part segmentationSAM2prompt-controllable segmentationmulti-view mask predictionLoRA fine-tuningnormal and point mapsclass-agnostic mIoUtextureless objects
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.

The reading

The paper tries to show that 3D part segmentation does not need text prompts, per-shape optimization, or full 3D labels. Instead, it reframes the task as multi-view 2D mask prediction: render a textureless object from twelve viewpoints, let a user click or box a part in one view, and let a SAM2 backbone—adapted with LoRA and normal-plus-point-map fusion—predict masks that are back-projected into consistent 3D parts. If the claim holds, interactive 2D prompts become a general control surface for fine-grained, object-level 3D understanding, with accuracy above existing optimization-based and feedforward pipelines and runtime near 30 seconds per object. The cost is that the training signal itself is generated by mesh-connectivity heuristics, so the method inherits whatever semantic gap exists between geometric components and human part concepts.

Core claim

GeoSAM2 claims that 3D part segmentation of a textureless object can be solved as promptable multi-view 2D mask prediction. The object is rendered from twelve canonical viewpoints into normal maps and point maps, arranged as a video sequence. A user clicks or draws a box on any view to indicate a target part, and a shared SAM2 backbone—kept mostly frozen but tuned with LoRA adapters and residual geometry fusion—predicts 2D masks for every view. These masks are back-projected to the surface with visibility-aware voting and post-processed by mesh connectivity, yielding consistent 3D part labels. The paper reports state-of-the-art class-agnostic mean IoU of 84.06 on PartObjaverse-Tiny and 74.42

What carries the argument

The engine is SAM2's video-mask loop repurposed as a 12-frame 'video' of normal and point maps. LoRA adapters (low-rank weight updates on every transformer attention layer) let the frozen backbone read geometric inputs; zero-initialized residual convolutions at each feature-pyramid level fuse point-map geometry into normal-map features without initially disturbing pretrained statistics; and a full-view memory bank—bootstrapped by repeating the prompted first frame—keeps masks consistent across large viewpoint changes. Back-projection with depth-based visibility checks and k-NN smoothing turns the per-view 2D masks into a final 3D labeling.

Load-bearing premise

The training labels are produced automatically by decomposing meshes along geometric connectivity and merging or splitting parts with heuristics; the method's accuracy on human-annotated benchmarks depends on those auto-labels matching human part semantics.

What would settle it

Retrain GeoSAM2 on the same roughly 4,700 objects but with human-labeled part annotations instead of the mesh-connectivity auto-labels, and evaluate on PartObjaverse-Tiny; if the margin over PartField collapses, the reported accuracy depends on the auto-labeling distribution rather than the prompt-controllable mechanism. Alternatively, fix prompts to a single user-chosen view without the best-of-four-views selection and measure the mIoU drop.

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

If this is right

  • A user can select an individual part—down to a patch like a leg or a beard—with a single click or box, and the selection propagates to a clean 3D mask, enabling fine-grained editing without text or scale tuning.
  • Since no per-shape optimization is required, inference completes in about 30 seconds per object, compared with minutes for optimization-based pipelines.
  • The method works on textureless meshes by fusing normal and point maps, so it applies to objects without RGB texture.
  • On PartObjaverse-Tiny and PartNetE, the paper reports higher class-agnostic mean IoU than Find3D, SAMPart3D, SAMesh, and PartField.
  • Because the backbone stays mostly frozen, future improvements to SAM2's mask quality can be absorbed by retraining only the small LoRA and fusion modules.

Where Pith is reading between the lines

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

  • The paper's evaluation selects the best of four orthographic input views; a fixed-view protocol would be a stricter test of the prompt-control claim and might narrow the reported margin over PartField.
  • Because the training signal comes from mesh connectivity, objects modeled as single continuous shells with no component boundaries may be systematically harder than artist-assembled objects; the method's ceiling is partly set by how CAD models are built.
  • The same frozen-SAM2 plus LoRA plus full-view memory recipe could transfer to other multi-view 3D lifting tasks, such as part completion or part-level editing, wherever a 2D prompt can name the region of interest.
Share X Bluesky LinkedIn Reddit HN

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 / 7 minor

Summary. GeoSAM2 casts 3D part segmentation of textureless objects as multi-view 2D mask prediction. It renders 12 normal and point maps from canonical viewpoints, feeds them through a frozen SAM2 backbone augmented with LoRA and zero-initialized residual fusion of geometric features, and uses a full-view memory bank with first-frame repetition to promote cross-view consistency. Predicted masks are back-projected and refined with mesh-aware post-processing. Training labels are generated automatically by mesh connectivity decomposition and heuristic merging/splitting. The paper claims state-of-the-art class-agnostic mIoU on PartObjaverse-Tiny and PartNetE, outperforming Find3D, SAMPart3D, SAMesh, and PartField, while running in about 30 seconds per object and requiring only 2D clicks or boxes.

Significance. If the empirical results are taken at face value, the paper makes a useful contribution: it replaces per-shape optimization and unintuitive global scale knobs with explicit 2D prompts, uses only automatically generated labels, and demonstrates consistent gains in the ablation study. The architecture design—LoRA on all attention projections, residual zero-conv fusion, and memory-bank redesign—is reasonable and well motivated. However, the central SOTA claim is currently not supported by the evaluation protocol. The comparison is weakened by the oracle selection of the best of four orthographic views using ground-truth mIoU, and the prompt-generation procedure for the quantitative benchmarks is not specified. These issues are load-bearing, not cosmetic, and must be fixed before the empirical claims can be accepted.

major comments (3)
  1. [§6.2, Tables 1 and 2] The paper states: "For our method, due to the object orientation uncertainty, we select 4 orthographics views as input respectively and adopt the highest mIoU." This is an oracle: the method is run four times with different starting views and the result with the highest ground-truth mIoU is reported. Baselines are evaluated once without an equivalent selection, so the comparison is not apples-to-apples and the reported margins (e.g., 84.06 vs. 79.18 on PartObjaverse-Tiny; 74.42 vs. 59.10 on PartNetE) are inflated. The paper should report a fixed-view protocol, the average over all four views, or a user-study-based protocol, and should clearly distinguish best-case from expected performance. In addition, the quantitative evaluation does not describe how 2D prompts are generated for each test object/part. If prompts are derived from ground-truth part locations, that is another form of test
  2. [§5.1 and Supplementary §8.1.2] The model is trained on labels produced by decomposing meshes according to connectivity and then merging/splitting parts with heuristics. The central claim of fine-grained, semantically meaningful part segmentation depends on these auto-labels aligning with human part semantics, but no quantitative validation of the auto-label quality is provided. I request a concrete check: compute agreement (e.g., mIoU or boundary accuracy) between the auto-decomposition and human ground-truth on a held-out subset, or show that a model trained on auto-labels transfers to human-annotated benchmarks without test-time fitting. Without such evidence, the semantic validity of the training signal remains an unsupported premise.
  3. [§6.2, Baselines] The text says both "For all methods, we follow their publicly released implementations and evaluate them on both datasets" and "The mIoU of the baselines is the same as reported in [15], as we adopt the same experimental setup." These statements are in tension. If baseline numbers are copied from [15], the paper must say so explicitly and verify that the metric, category grouping, and evaluation splits are identical. If they were re-run, the paper should report the actual numbers and any variation. This is needed for reproducibility and for a fair comparison, especially because the oracle view selection is applied only to GeoSAM2.
minor comments (7)
  1. [§5.2] Typo: "repectively" should be "respectively."
  2. [§6.2] Typo: "generatie" should be "generate."
  3. [§8] Typo: "Supplmentary" should be "Supplementary."
  4. [References] The SAMesh reference is cited inconsistently as [26], [27], and [28] with different years and titles. Please unify the citation and verify the bibliographic entry. Also, PartNetE is referenced as [14], which appears to be PartSLIP's arXiv entry; please cite the actual dataset source.
  5. [Tables 1 and 2] The table formatting is difficult to read: several category columns appear shifted or duplicated, and the per-category and average columns are not cleanly aligned. Please regenerate the tables with clear column separation.
  6. [§6.2] The sentence "we also automatically segments the unlabeled area of the view which lies opposite the input view" is vague. Please describe the mechanism and specify whether it uses any ground-truth information or is a fixed heuristic.
  7. [§6.4, Table 3] The ablation tables (Tables 3–5) report per-category results in the supplement, but the main text does not mention that the "w/o feature fusion" variant uses a 1x1 convolution while the full method uses a 3x3 kernel. This detail should be in the main text or clearly cross-referenced.

Circularity Check

1 steps flagged

Oracle 4-view mIoU selection turns the reported SOTA numbers into test-set-fitted upper bounds; the central comparison is partially circular.

specific steps
  1. fitted input called prediction [Section 6.2 (Comparision), evaluation protocol]
    "For our method, due to the object orientation uncertainty, we select 4 orthographics views as input respectively and adopt the highest mIoU."

    The reported mIoU for GeoSAM2 is not the output of a fixed prediction rule; it is the maximum mIoU over four input orientations, where the selection is made using the ground-truth mIoU itself. Thus the test labels are used to choose a method configuration, and the headline numbers in Tables 1 and 2 (84.06 and 74.42) are oracle upper bounds rather than expected interactive performance. Because baselines are evaluated without an equivalent oracle selection, the SOTA margins are inflated by construction. The claim 'we achieve state-of-the-art class-agnostic performance' therefore reduces, at the evaluation step, to a test-set-fitted maximum rather than a genuine prediction.

full rationale

The core learning pipeline is not circular: training labels are generated from mesh connectivity (§5.1, §8.1.2), separate from the human-annotated PartObjaverse-Tiny/PartNetE evaluation sets, and the LoRA/fusion/memory components are tested against external benchmarks. The auto-labeling concern raised by the reader is a validity issue (do connectivity-derived parts match human semantics?), not a circularity issue, because the evaluation uses independent human labels. No load-bearing self-citation or imported-uniqueness step was found; SAMPart3D and HoloPart citations are contextual or application-level. However, the central SOTA claim is partially circular due to §6.2's oracle view selection: the metric used to report success also selects which of four runs counts as 'the method'. This makes the reported numbers best-case upper bounds and the comparison to single-run baselines asymmetric. Score 6 reflects partial circularity in the evaluation claim, not in the architectural derivation.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central claim rests on a set of hand-chosen hyperparameters and domain assumptions. The most significant free parameter is the evaluation protocol's oracle view selection, which fits the reported metric to the test set. The training data itself is derived from a heuristic mesh decomposition, an unverified assumption about semantic alignment with human annotations.

free parameters (7)
  • Evaluation view selection (oracle) = 4 orthographic views, highest mIoU selected
    Section 6.2: the authors select the view that yields the highest mIoU using ground truth, inflating reported performance.
  • Number of views and elevations = 12 views, elevations 25/0/-25 degrees
    Chosen by hand in Section 5.3; affects coverage and memory.
  • LoRA rank = 4
    Set in Section 6.1; controls adaptation capacity.
  • Post-processing area threshold = 0.01 of total area
    Set in Section 5.4; removes small components.
  • Training epochs = 50
    Set in Section 6.1.
  • Learning rate = 5e-5
    Set in Section 6.1.
  • Part merging threshold = 15 parts
    Section 5.1: objects with more than 15 parts are recomposed; heuristic choice.
axioms (5)
  • domain assumption Mesh connectivity decomposition yields semantically meaningful part labels
    Section 5.1 and Supplementary 8.1.2: training labels come from connectivity splitting and merging heuristics, not human annotation.
  • domain assumption SAM2's pretrained RGB priors transfer to normal and point maps after LoRA adaptation
    Section 5.2: assumes LoRA steers SAM2 to geometric modalities while preserving priors.
  • domain assumption 12 canonical views provide sufficient geometric coverage
    Section 5.3: views are evenly spaced in azimuth and three elevations; heavy occlusion is acknowledged as a limitation in Section 7.1.
  • domain assumption Full-view memory retention and frame repetition enable cross-view consistency
    Section 5.3: redesign of SAM2's memory, validated only by the paper's experiments.
  • standard math Standard pinhole rendering and back-projection equations
    Sections 5 and 8.1.1: depth-based visibility checking and inverse projection.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GeoSAM2: Unleashing the Power of SAM2 for 3D Part Segmentation." pith.science (2026). https://pith.science/paper/6XDW4JQB

@misc{pith2026250814036,
  author       = {Pith},
  title        = {Pith review of: GeoSAM2: Unleashing the Power of SAM2 for 3D Part Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6XDW4JQB}},
  note         = {Machine review of arXiv:2508.14036}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce GeoSAM2, a prompt-controllable framework for 3D part segmentation that casts the task as multi-view 2D mask prediction. Given a textureless object, we render normal and point maps from predefined viewpoints and accept simple 2D prompts - clicks or boxes - to guide part selection. These prompts are processed by a shared SAM2 backbone augmented with LoRA and residual geometry fusion, enabling view-specific reasoning while preserving pretrained priors. The predicted masks are back-projected to the object and aggregated across views. Our method enables fine-grained, part-specific control without requiring text prompts, per-shape optimization, or full 3D labels. In contrast to global clustering or scale-based methods, prompts are explicit, spatially grounded, and interpretable. We achieve state-of-the-art class-agnostic performance on PartObjaverse-Tiny and PartNetE, outperforming both slow optimization-based pipelines and fast but coarse feedforward approaches. Our results highlight a new paradigm: aligning the paradigm of 3D segmentation with SAM2, leveraging interactive 2D inputs to unlock controllability and precision in object-level part understanding.

Figures

Figures reproduced from arXiv: 2508.14036 by Ding Liang, Jingxiang Sun, Ken Deng, Xihui Liu, Yan-Pei Cao, Yebin Liu, Yunhan Yang.

Figure 1
Figure 1. Figure 1: We introduce GeoSAM2, a scalpel-precision 3D shape segmentation method that enables arbitrary-detail segmentation through [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Our pipeline renders 12-view normal and point maps of an object, arranged into an inverse-clockwise video sequence. Users can [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of class-agnostic segmentation on the PartObjaverse-Tiny dataset [36]. The baselines include Find3D [17], SAMesh [28], and SAMPart3D [36], PartField [15]. Each color represents a different part. More results can be found in supplementary material [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of hierarchical segmentation results of gen [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of ablation study on PartObjaverse-Tiny. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study of frame repetition . 6.3. Generalization to Generative Model To evaluate the generalizability of our method, we extend our method to the generation results. Because the 3D gen￾eration model still can not generate handcrafted-like geom￾etry with sharp edges and reasonable structures, segment those generation results is challenging. We perform hierarchical segmentation on 3D models generated … view at source ↗
Figure 8
Figure 8. Figure 8: Example of segmenting a 3D part by given a single 2D [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Examples of training data. Each color represents a dif [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Illustrations of w/o feature fusion and our full method in abla￾tion study. ants used [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: More comparison results on PartObjaverse-Tiny [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: More comparison results on PartObjaverse-Tiny [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: More hierarchical segmentation visualization. 3D models are segmented hierarchically, starting with the finest level of detail at [PITH_FULL_IMAGE:figures/full_fig_p013_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: More hierarchical segmentation visualization. 3D models are segmented hierarchically, starting with the finest level of detail at [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: More hierarchical segmentation visualization. 3D models are segmented hierarchically, starting with the finest level of detail at [PITH_FULL_IMAGE:figures/full_fig_p015_15.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Part-Level 3D Gaussian Vehicle Generation with Joint and Hinge Axis Estimation

    cs.AI 2026-04 unverdicted novelty 6.0

    A new framework generates part-level animatable 3D Gaussian vehicles from images by adding modules for exclusive part ownership and kinematic joint/axis prediction.

Reference graph

Works this paper leans on

42 extracted references · 30 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    Satr: Zero-shot semantic segmentation of 3d shapes, 2023

    Ahmed Abdelreheem, Ivan Skorokhodov, Maks Ovsjanikov, and Peter Wonka. Satr: Zero-shot semantic segmentation of 3d shapes, 2023. 3

    work page 2023

  2. [2]

    Seg- ment anything in 3d with radiance fields, 2024

    Jiazhong Cen, Jiemin Fang, Zanwei Zhou, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. Seg- ment anything in 3d with radiance fields, 2024. 3

    work page 2024

  3. [3]

    A benchmark for 3D mesh segmentation

    Xiaobai Chen, Aleksey Golovinskiy, and Thomas Funkhouser. A benchmark for 3D mesh segmentation. ACM Transactions on Graphics (Proc. SIGGRAPH) , 28(3),

    work page

  4. [4]

    3d part seg- mentation via geometric aggregation of 2d visual features

    Marco Garosi, Riccardo Tedoldi, Davide Boscaini, Massi- miliano Mancini, Nicu Sebe, and Fabio Poiesi. 3d part seg- mentation via geometric aggregation of 2d visual features. In 2025 IEEE/CVF Winter Conference on Applications of Com- puter Vision (WACV), pages 3257–3267. IEEE, 2025. 3

    work page 2025

  5. [5]

    Meshcnn: a network with an edge

    Rana Hanocka, Amir Hertz, Noa Fish, Raja Giryes, Shachar Fleishman, and Daniel Cohen-Or. Meshcnn: a network with an edge. ACM Transactions on Graphics, 38(4):1–12, 2019. 3

    work page 2019

  6. [6]

    Segment3D: Learning Fine-Grained Class-Agnostic 3D Segmentation without Manual Labels

    Rui Huang, Songyou Peng, Ayca Takmaz, Federico Tombari, Marc Pollefeys, Shiji Song, Gao Huang, and Francis En- gelmann. Segment3d: Learning fine-grained class-agnostic 3d segmentation without manual labels. arXiv preprint arXiv:2312.17232, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything, 2023. 3

    2023

  8. [9]

    Grounded language-image pre-training

    Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jian- wei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 3

    work page 2022

  9. [10]

    Laplacian mesh transformer: Dual attention and topology aware net- work for 3d mesh classification and segmentation

    Xiao-Juan Li, Jie Yang, and Fang-Lue Zhang. Laplacian mesh transformer: Dual attention and topology aware net- work for 3d mesh classification and segmentation. In Com- puter Vision – ECCV 2022 , pages 541–560, Cham, 2022. Springer Nature Switzerland. 3

    work page 2022

  10. [11]

    Pointcnn: Convolution on x-transformed points

    Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. In NeurIPS, 2018. 2

    work page 2018

  11. [12]

    Triposg: High- fidelity 3d shape synthesis using large-scale rectified flow models, 2025

    Yangguang Li, Zi-Xin Zou, Zexiang Liu, Dehu Wang, Yuan Liang, Zhipeng Yu, Xingchao Liu, Yuan-Chen Guo, Ding Liang, Wanli Ouyang, and Yan-Pei Cao. Triposg: High- fidelity 3d shape synthesis using large-scale rectified flow models, 2025. 7

    work page 2025

  12. [13]

    Partslip: Low-shot part seg- mentation for 3d point clouds via pretrained image-language models

    Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part seg- mentation for 3d point clouds via pretrained image-language models. In CVPR, 2023. 3

    work page 2023

  13. [14]

    Partslip: Low-shot part seg- mentation for 3d point clouds via pretrained image-language models, 2023

    Minghua Liu, Yinhao Zhu, Hong Cai, Shizhong Han, Zhan Ling, Fatih Porikli, and Hao Su. Partslip: Low-shot part seg- mentation for 3d point clouds via pretrained image-language models, 2023. 6

    work page 2023

  14. [15]

    Partfield: Learn- ing 3d feature fields for part segmentation and beyond.arXiv preprint arXiv:2504.11451, 2025

    Minghua Liu, Mikaela Angelina Uy, Donglai Xiang, Hao Su, Sanja Fidler, Nicholas Sharp, and Jun Gao. Partfield: Learn- ing 3d feature fields for part segmentation and beyond.arXiv preprint arXiv:2504.11451, 2025. 6, 7

    Pith/arXiv arXiv 2025

  15. [16]

    SANeRF-HQ: Segment Anything for NeRF in High Quality

    Yichen Liu, Benran Hu, Chi-Keung Tang, and Yu-Wing Tai. Sanerf-hq: Segment anything for nerf in high quality. arXiv preprint arXiv:2312.01531, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  16. [17]

    Find any part in 3d, 2025

    Ziqi Ma, Yisong Yue, and Georgia Gkioxari. Find any part in 3d, 2025. 2, 6

    work page 2025

  17. [18]

    Partnet: A large- scale benchmark for fine-grained and hierarchical part-level 3d object understanding

    Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A large- scale benchmark for fine-grained and hierarchical part-level 3d object understanding. In CVPR, 2019. 3

    work page 2019

  18. [19]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth ´ee Darcet, Th ´eo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 3

    Pith/arXiv arXiv 2023

  19. [20]

    Better call sal: Towards learning to segment anything in lidar.arXiv preprint arXiv:2403.13129, 2024

    Aljo ˇsa O ˇsep, Tim Meinhardt, Francesco Ferroni, Neehar Peri, Deva Ramanan, and Laura Leal-Taix ´e. Better call sal: Towards learning to segment anything in lidar.arXiv preprint arXiv:2403.13129, 2024. 3

    Pith/arXiv arXiv 2024

  20. [21]

    Openscene: 3d scene understanding with open vocabularies

    Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 815–824, 2023. 3

    work page 2023

  21. [22]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 3

    work page 2017

  22. [23]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In ICML, 2021. 3

    2021

  23. [24]

    Neural shape diameter function for efficient mesh segmentation

    Bruno Roy. Neural shape diameter function for efficient mesh segmentation. In ACM SIGGRAPH 2023 Posters . ACM, 2023. 3

    work page 2023

  24. [25]

    Consis- tent mesh partitioning and skeletonisation using the shape diameter function

    Lior Shapira, Ariel Shamir, and Daniel Cohen-Or. Consis- tent mesh partitioning and skeletonisation using the shape diameter function. The Visual Computer, 24:249–259, 2008. 3

    work page 2008

  25. [26]

    Segment any mesh: Zero-shot mesh part segmentation via lifting segment anything 2 to 3d

    George Tang, William Zhao, Logan Ford, David Benhaim, and Paul Zhang. Segment any mesh: Zero-shot mesh part segmentation via lifting segment anything 2 to 3d. arXiv:2408.13679, 2024. 2

    Pith/arXiv arXiv 2024

  26. [27]

    Segment any mesh: Zero-shot mesh part segmentation via lifting segment anything 2 to 3d, 2024

    George Tang, William Zhao, Logan Ford, David Benhaim, and Paul Zhang. Segment any mesh: Zero-shot mesh part segmentation via lifting segment anything 2 to 3d, 2024. 6

    work page 2024

  27. [28]

    Segment any mesh, 2025

    George Tang, William Zhao, Logan Ford, David Benhaim, and Paul Zhang. Segment any mesh, 2025. 6

    work page 2025

  28. [29]

    Generating part-aware editable 3d shapes without 3d supervision

    Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan, Jeong Joon Park, Mikaela Angelina Uy, Ioannis Emiris, Yan- nis Avrithis, and Leonidas Guibas. Generating part-aware editable 3d shapes without 3d supervision. In CVPR, 2023. 2 16

    work page 2023

  29. [30]

    PartDistill: 3D Shape Part Segmentation by Vision-Language Model Distillation

    Ardian Umam, Cheng-Kun Yang, Min-Hung Chen, Jen- Hui Chuang, and Yen-Yu Lin. Partdistill: 3d shape part segmentation by vision-language model distillation. arXiv:2312.04016, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [31]

    Open-vocabulary part-based grasping

    Tjeard van Oort, Dimity Miller, Will N Browne, Nico- las Marticorena, Jesse Haviland, and Niko Suenderhauf. Open-vocabulary part-based grasping. arXiv preprint arXiv:2406.05951, 2024. 2

    arXiv 2024

  31. [32]

    Coseg: Cognitively inspired unsupervised generic event segmenta- tion

    Xiao Wang, Jingen Liu, Tao Mei, and Jiebo Luo. Coseg: Cognitively inspired unsupervised generic event segmenta- tion. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3

    work page 2023

  32. [33]

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

    Mutian Xu, Xingyilang Yin, Lingteng Qiu, Yang Liu, Xin Tong, and Xiaoguang Han. Sampro3d: Locating sam prompts in 3d for zero-shot scene segmentation. arXiv preprint arXiv:2311.17707, 2023. 3

    Pith/arXiv arXiv 2023

  33. [34]

    ZeroPS: High-quality Cross-modal Knowledge Transfer for Zero-Shot 3D Part Segmentation

    Yuheng Xue, Nenglun Chen, Jun Liu, and Wenyun Sun. Ze- rops: High-quality cross-modal knowledge transfer for zero- shot 3d part segmentation. arXiv:2311.14262, 2023. 3

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  34. [35]

    Sam3d: Segment anything in 3d scenes

    Yunhan Yang, Xiaoyang Wu, Tong He, Hengshuang Zhao, and Xihui Liu. Sam3d: Segment anything in 3d scenes. arXiv:2306.03908, 2023. 2, 3

    Pith/arXiv arXiv 2023

  35. [36]

    Sampart3d: Segment any part in 3d objects

    Yunhan Yang, Yukun Huang, Yuan-Chen Guo, Liangjun Lu, Xiaoyang Wu, Edmund Y Lam, Yan-Pei Cao, and Xihui Liu. Sampart3d: Segment any part in 3d objects. arXiv preprint arXiv:2411.07184, 2024. 3, 6

    Pith/arXiv arXiv 2024

  36. [37]

    Holopart: Generative 3d part amodal segmentation, 2025

    Yunhan Yang, Yuan-Chen Guo, Yukun Huang, Zi-Xin Zou, Zhipeng Yu, Yangguang Li, Yan-Pei Cao, and Xihui Liu. Holopart: Generative 3d part amodal segmentation, 2025. 1, 2

    work page 2025

  37. [38]

    Point transformer

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 3

    work page 2021

  38. [39]

    Meshsegmenter: Zero-shot mesh semantic segmentation via texture synthesis

    Ziming Zhong, Yanyu Xu, Jing Li, Jiale Xu, Zhengxin Li, Chaohui Yu, and Shenghua Gao. Meshsegmenter: Zero-shot mesh semantic segmentation via texture synthesis. InECCV. Springer, 2024. 3

    work page 2024

  39. [40]

    Serf: Fine-grained interactive 3d segmentation and editing with radiance fields, 2024

    Kaichen Zhou, Lanqing Hong, Enze Xie, Yongxin Yang, Zhenguo Li, and Wei Zhang. Serf: Fine-grained interactive 3d segmentation and editing with radiance fields, 2024. 3

    work page 2024

  40. [41]

    Partslip++: Enhancing low-shot 3d part segmentation via multi-view instance segmentation and maximum likelihood estimation

    Yuchen Zhou, Jiayuan Gu, Xuanlin Li, Minghua Liu, Yun- hao Fang, and Hao Su. Partslip++: Enhancing low-shot 3d part segmentation via multi-view instance segmentation and maximum likelihood estimation. arXiv:2312.03015, 2023. 3

    Pith/arXiv arXiv 2023

  41. [42]

    Point-sam: Promptable 3d segmentation model for point clouds

    Yuchen Zhou, Jiayuan Gu, Tung Yen Chiang, Fanbo Xiang, and Hao Su. Point-sam: Promptable 3d segmentation model for point clouds. arXiv preprint arXiv:2406.17741, 2024. 3

    Pith/arXiv arXiv 2024

  42. [43]

    Point- clip v2: Prompting clip and gpt for powerful 3d open-world learning

    Xiangyang Zhu, Renrui Zhang, Bowei He, Ziyu Guo, Ziyao Zeng, Zipeng Qin, Shanghang Zhang, and Peng Gao. Point- clip v2: Prompting clip and gpt for powerful 3d open-world learning. In ICCV, 2023. 3 17

    work page 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.