Pith. sign in

REVIEW 3 major objections 5 minor 85 references

UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read UniPre3D claims one pre-training objective works for object- and scene-level point clouds, using Gaussian splatting to render predicted primitives into images.

desk verdict Real novelty in the rendering-based pre-training recipe, but the 'unified/any scale' headline is not supported by the paper's own two-protocol design; still worth refereeing. read the letter →

arxiv 2506.09952 v1 pith:V6UGSNUL submitted 2025-06-11 cs.CV cs.AI

Ziyi Wang , Yanran Zhang , Jie Zhou , Jiwen Lu This is my paper
classification cs.CVcs.AI
keywords unifiedpre-training3DpointcloudGaussianSplattingcross-modalfusiongenerativesceneunderstandingobjectclassification
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

Existing pre-training methods for 3D point clouds split into two camps: masked autoencoding works on small object-level point clouds, contrastive learning works on large scene-level point clouds, and neither transfers to the other scale. This paper introduces UniPre3D, a single generative pre-training method that claims to work on both scales and across point cloud architectures of different kinds. The idea is to make the 3D model predict a set of Gaussian primitives from the input point cloud, render them into images with differentiable Gaussian splatting, and supervise the rendered images against ground-truth views with pixel-level loss, while using features from a pre-trained 2D image model to supply color and texture cues. If the claim holds, one pre-training pipeline could replace the scale-specific paradigms that currently fragment 3D representation learning, and pre-training could be organized around a rendering objective regardless of whether the input is a single object or a full indoor scene.

What carries the argument

The load-bearing mechanism is 3D Gaussian Splatting: a lightweight predictor head outputs, for each point, a Gaussian primitive with position offset, opacity, scale, rotation quaternion, and spherical harmonics coefficients, and a differentiable renderer projects these primitives into images, giving full-resolution pixel-level supervision and end-to-end gradient flow. The second mechanism is scale-adaptive cross-modal fusion: 2D features extracted by a pre-trained image model are projected into 3D space and combined with the backbone features, either by concatenating decoder features with per-point 2D features for objects, or by back-projecting pixels into a pseudo point cloud that is voxel-averaged together with the encoder's first layer for scenes. This combination lets the pre-training task be hard enough to teach geometry but not so hard that the backbone fails to learn.

What would settle it

Pre-train the object-level model on ShapeNet with the 2D branch completely removed and fine-tune on ScanObjectNN PB T50 RS: if the accuracy gain over no pre-training is statistically indistinguishable whether or not the image features were used, then the cross-modal fusion is not load-bearing. Likewise, pre-train the scene-level model on a depth-free version of ScanNet where point fusion cannot be applied: if the ScanNet200 mIoU gain over no pre-training collapses, then the unified claim holds only when ground-truth depth is available.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that predicting Gaussian primitives and rendering them to images is a universal pre-training objective for 3D point clouds: it can be applied seamlessly to point clouds of any scale and to 3D models of any architecture, and it outperforms scale-specific pre-training methods on most benchmarks. To regulate task difficulty, the method fuses features from a pre-trained image model with the backbone's 3D features using two different strategies—feature fusion at the decoder for object-level data and point fusion (back-projecting pixels into a denser voxelized pseudo point cloud) for scene-level data. The paper validates the claim on object classification (ScanObjectNN), part segmentation (ShapeNetPart), scene semantic segmentation (ScanNet20, ScanNet200, S3DIS), instance segmentation, and scene object detection, using Transformer, MLP, state-space, sparse-convolution, and point-transformer backbones.

Load-bearing premise

The whole method rests on the premise that training a 3D network to render pixels—helped by 2D features it will not have at test time—teaches geometric structure that survives when the image branch is removed and the backbone is fine-tuned for recognition.

Editorial extensions

If this is right

  • A single pre-training recipe replaces the object/scene split: the same rendering objective drives both object classification and scene segmentation pre-training.
  • The method transfers across architectures, improving standard Transformers, PointMLP, Mamba3D, PointCloudMamba, SparseUNet, and PointTransformerV3, so the gain does not depend on one backbone design.
  • Full-image pixel supervision from 3DGS is feasible where NeRF-based counterparts render only a subset of pixels, and the pre-training runs roughly twice as fast as PonderV2 on scene-level experiments.
  • Gains are strongest on the long-tail ScanNet200 benchmark, suggesting the pre-training particularly helps on rare and small-object classes.
  • The ablation shows fusion depth must be chosen: fusing 2D features at every decoder layer raises rendering PSNR but lowers downstream accuracy, so the cross-modal signal needs to be gated.

Reading between the lines

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

  • The paper's two fusion strategies are chosen by data scale; an obvious extension the paper does not explore is automatically selecting the fusion point from the point density or the availability of depth, which would remove the manual choice.
  • If rendering quality is a proxy for pre-training effectiveness, the PSNR-versus-downstream-accuracy inversion in the ablations suggests that 2D features can overfit the rendering task; a testable extension is to anneal the 2D feature contribution during pre-training instead of freezing it.
  • The idea of using the image domain as an intermediary to absorb scale differences could generalize beyond point clouds to other irregular 3D representations such as depth maps, meshes, or implicit fields, wherever a differentiable projector exists.
  • Because the scene-level point fusion requires ground-truth depth, the method's universality on real-world scans without depth would depend on predicted depth; testing UniPre3D with monocular depth estimation would bound how far the unified claim extends.
Share X Bluesky LinkedIn Reddit HN

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 UniPre3D, a pre-training framework for 3D point cloud models that predicts Gaussian primitives from point-cloud features and renders images via differentiable Gaussian splatting, supervised by pixel-wise MSE against ground-truth reference views. To modulate task difficulty, the method fuses features from a pre-trained 2D image model into the 3D backbone using one of two scale-specific mechanisms: object-level feature fusion (projecting 3D points into image space, Eqs. 3-6) and scene-level point fusion (back-projecting depth maps into a pseudo point cloud, Eqs. 7-9). The authors claim this is the first unified pre-training approach that can be seamlessly applied to point clouds of any scale and to 3D models of any architecture. Experiments cover object classification on ScanObjectNN, part segmentation on ShapeNetPart, semantic and instance segmentation on ScanNet20/200 and S3DIS, and detection on ScanNet20, using Transformer, PointMLP, PointCloudMamba, Mamba3D, SparseUNet, and PointTransformerV3 backbones. Ablations study the fusion layer and fusion strategy. The paper includes code and supplementary detection and view-selection experiments.

Significance. If validated, the core idea—using 3D Gaussian splatting as a unified generative pre-training objective that works for both object-level and scene-level point clouds—is a meaningful contribution to 3D representation learning. The paper is honest in reporting limitations and provides extensive experiments across diverse tasks and backbones, including strong results on ScanObjectNN PB T50 RS (87.93 with a standard Transformer) and ScanNet200 semantic segmentation (33.0 mIoU with SparseUNet, 36.0 with PTv3). The release of code is also a strength. However, the central 'unified and seamless' claim is substantially qualified by the paper's own Section 4.4, which concedes the need for manual fusion-strategy selection, and by the experimental protocol that uses two distinct, non-interchangeable recipes. The empirical evidence would also be stronger with repeated runs and error bars, given that several headline gains are small (e.g., Table 3, ScanNet20 PTv3: 77.45→77.63). Overall, the contribution is defensible as a broadly applicable pre-training framework, but the presentation overstates the degree of unification.

major comments (3)
  1. [§3.3 and §4.4] The abstract and introduction claim that UniPre3D can be 'seamlessly applied' to point clouds of any scale, but Section 3.3 defines two separate mechanisms—object-level feature fusion (Eqs. 3-6) and scene-level point fusion (Eqs. 7-9)—that require different supervision (no depth vs. ground-truth depth) and are manually selected per scale. Table 6 shows the two recipes are not interchangeable: using object-style feature fusion on scenes yields lower mIoU than point fusion (ScanNet20 75.7 vs 75.8, ScanNet200 32.3 vs 33.0, S3DIS 70.9 vs 71.5). Section 4.4 explicitly states that 'the manual fusion strategy selection further limits unification.' This directly weakens the headline contribution: as presented, the method is a family of two scale-specific pre-training protocols, not a single unified algorithm. I recommend reframing the claim to describe a unified pre-training framework with scale-adaptive components, or, if the 'seamless' claim is to be retained, providing an automatic selection mechanism or evidence that the choice can be made without dataset-specific knowledge.
  2. [§4.2, Tables 1-4 and Table 7] All downstream results are reported as single numbers with no error bars or repeated-seed statistics. Several improvements over strong baselines are small enough that they could fall within run-to-run variance, e.g., ScanNet20 PTv3 (77.45→77.63 in Table 3), ScanNet200 PTv3 (35.2→36.0), and S3DIS SparseUNet (71.5 vs 70.2 for PonderV2† and 71.9 for PPT). Without variance estimates, the claim that UniPre3D 'consistently outperforms' prior methods is not fully supported. I request adding at least three seeds with mean and standard deviation for the main comparisons, or, if computation is prohibitive, explicitly stating that results are single-run and flagging which differences are likely not significant.
  3. [§4.2.2 (Tables 3)] Results for PTv3 on S3DIS are omitted, with the explanation that the official implementation requires disabling flash-attention, exceeding GPU memory. This omission weakens the claim of universal effectiveness across both scenes and advanced backbones, especially since PTv3 is the strongest scene backbone tested. At minimum, the omitted cell should be clearly marked as 'not evaluated' in the table (it currently appears as an em dash), and the authors should either provide the result using a memory-efficient implementation or discuss in the limitations why the method cannot be evaluated on this combination.
minor comments (5)
  1. [§3.3, Eqs. (3)-(5)] The notation for the camera extrinsic matrix V and the depth variable d is introduced only in the surrounding text rather than in the equations; consider defining all symbols directly after the equations to improve readability.
  2. [§4.1] The phrase 'This requires one NVIDIA 3090Ti GPU' is informal; suggest 'a single NVIDIA 3090Ti GPU' and, for consistency, specify the GPU memory (24GB) for the scene-level setup as well.
  3. [References] Reference [46] (Rozenberszki et al., Language-grounding indoor 3D semantic segmentation in the wild) is duplicated as reference [47]; also, the title of reference [77] contains a typo: 'Point Could Mamba' should be 'Point Cloud Mamba'.
  4. [§4.3 and Tables 5-6] The PSNR values are reported to one decimal place and differ only slightly across fusion variants (e.g., 16.6 to 16.8 in Table 6); please clarify whether these differences are stable across training runs, since the fine-tuning mIoU differences are also small.
  5. [§4.2.1 and Table 2] In the part segmentation results, UniPre3D achieves 86.8 mIoU_I, which is 0.1 lower than TAP's 86.9; the text says 'competitive results with TAP on mIoU_I,' which is accurate, but the phrasing could be more explicit about this small gap.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the pre-training objective and reported downstream gains are independently defined, with the only flagged limitation being a scope issue, not a circular derivation.

full rationale

No circular step is present. The pre-training objective (Eq. 10) is a pixel-wise MSE between rendered and ground-truth images, while the reported claims are downstream fine-tuning accuracies on ScanObjectNN, ShapeNetPart, ScanNet20/200, S3DIS, and ScanNet detection — external benchmarks not defined by the pre-training loss. The paper's design choices (fusion-layer location, point vs. feature fusion, reference-view counts) are tuned via ablations in Tables 5, 6, and 8, which is standard empirical model selection, not a fitted parameter being renamed as a prediction. The self-identified limitation in Section 4.4 that 'the manual fusion strategy selection further limits unification' is a scope/correctness concern about the 'seamlessly applied' claim, not a circular derivation: no equation is defined in terms of the conclusion, and no load-bearing claim is justified solely by a self-citation. The comparison protocols follow prior work (single-dataset unsupervised pre-training), and reproduced PonderV2 numbers are marked as such. Therefore, under the requested standard, the correct finding is no significant circularity.

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

No new physical or architectural entity is invented. Gaussian primitives are taken from prior 3DGS work. The central claim rests on four assumptions about the transferability of rendering-based pre-training and on several manually tuned hyperparameters (loss weights, view counts, fusion strategy).

free parameters (4)
  • object-level loss weights omega_fg, omega_bg = omega_fg=4, omega_bg=1
    Set by hand to balance foreground/background pixel loss (Section 3.4, Section 4.1); downstream accuracy varies with this choice, but no sensitivity analysis is reported.
  • reference view count = 1 (object), 8 (scene)
    Chosen by ablation (Supplementary A.2, Table 8); the task complexity depends on this and is part of the scale-adaptive design.
  • perspective gap threshold = fewer than 5 images
    Scene pre-training restricts rendered views near reference views (Section 4.1, Supplementary Table 8); ablations show this affects mIoU by about 0.7.
  • fusion strategy = feature fusion (object), point fusion (scene)
    Manual selection per scale (Section 3.3, Table 6); acknowledged as a limitation in Section 4.4 and undermines the "seamless/self-adaptive" claim.
assumptions (4)
  • domain assumption 3DGS rendering loss is a useful pre-training target for downstream 3D perception.
    The whole method assumes pixel-level rendering supervision transfers to classification/segmentation accuracy; Section 3.2 states this as design insight with no theoretical guarantee.
  • domain assumption Pre-trained image features supply well-established texture knowledge that helps rather than hurts geometric pre-training.
    Invoked in Section 3.3; the paper's own ablations (Section 4.3) show too much 2D fusion can hurt fine-tuning, so the direction and degree of benefit is empirical.
  • domain assumption Point cloud information density aligns with projected image information density across scales.
    Used in Section 3.2 to justify why Gaussian rendering provides adaptive task difficulty.
  • domain assumption Standard benchmark protocols (ScanObjectNN, ShapeNetPart, ScanNet, S3DIS) measure the claimed universal effectiveness.
    The paper uses these benchmarks as the sole evidence for universal effectiveness; this assumes the chosen tasks and splits are representative of 3D perception needs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting." pith.science (2026). https://pith.science/paper/V6UGSNUL

@misc{pith2026250609952,
  author       = {Pith},
  title        = {Pith review of: UniPre3D: Unified Pre-training of 3D Point Cloud Models with Cross-Modal Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6UGSNUL}},
  note         = {Machine review of arXiv:2506.09952}
}
read the original abstract

The scale diversity of point cloud data presents significant challenges in developing unified representation learning techniques for 3D vision. Currently, there are few unified 3D models, and no existing pre-training method is equally effective for both object- and scene-level point clouds. In this paper, we introduce UniPre3D, the first unified pre-training method that can be seamlessly applied to point clouds of any scale and 3D models of any architecture. Our approach predicts Gaussian primitives as the pre-training task and employs differentiable Gaussian splatting to render images, enabling precise pixel-level supervision and end-to-end optimization. To further regulate the complexity of the pre-training task and direct the model's focus toward geometric structures, we integrate 2D features from pre-trained image models to incorporate well-established texture knowledge. We validate the universal effectiveness of our proposed method through extensive experiments across a variety of object- and scene-level tasks, using diverse point cloud models as backbones. Code is available at https://github.com/wangzy22/UniPre3D.

Figures

Figures reproduced from arXiv: 2506.09952 by the authors.

Figure 1
Figure 1. Pre-training paradigm comparison. Existing object￾level pre-training methods usually follow a generative masked auto-encoding (MAE) paradigm. Their scene-level counterparts mostly leverage the contrastive learning paradigm. We propose a unified pre-training method that is applicable and effective to both object- and scene-level point clouds and models. tain hundreds of times more points than object-level point cloud… view at source ↗
Figure 2
Figure 2. UniPre3D pre-training pipeline. Our proposed pre￾training task involves predicting Gaussian parameters from the in￾put point cloud. The 3D backbone network is expected to extract representative features, and 3D Gaussian splatting is implemented to render images for direct supervision. To incorporate additional texture information and adjust task complexity, we introduce a pre￾trained image model and propose a scale-… view at source ↗
Figure 3
Figure 3. Visualization of UniPre3D pre-training outputs. The first row presents the input point clouds, followed by the reference view images in the second row. The third row displays the rendered images, which are supervised by the ground truth images shown in the fourth row. In the rightmost column, we illustrate a schematic diagram of the view selection principle for both object- and scene-level samples. also implemented … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of UniPre3D pre-training outputs on object-level experiments. The first column presents the input point clouds, followed by the reference view images in the second column. The remaining rows display the rendered images. Point Cloud Reference Images Render…
Figure 5
Figure 5. Figure 5: Visualization of UniPre3D pre-training outputs on scene-level experiments. The first column presents the input point clouds, followed by the reference view images in the second and third columns. The remaining columns display the rendered images (upper rows) and their …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

85 extracted references · 64 canonical work pages

  1. [1]

    3d semantic parsing of large-scale indoor spaces

    Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 6

    2016

  2. [2]

    Shapenet: An information-rich 3d model repository

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5, 7

    arXiv 2015

  3. [3]

    pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction

    David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR,

  4. [4]

    Decoupled local aggregation for point cloud learning

    Binjie Chen, Yunzhou Xia, Yu Zang, Cheng Wang, and Jonathan Li. Decoupled local aggregation for point cloud learning. arXiv preprint arXiv:2308.16532, 2023. 2

    arXiv 2023

  5. [5]

    Pointgpt: Auto-regressively generative pre- training from point clouds

    Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pre- training from point clouds. NeurIPS, 2024. 2

    2024

  6. [6]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2

    2020

  7. [7]

    Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images

    Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. In ECCV, 2025. 3

    2025

  8. [8]

    Unit3d: A unified transformer for 3d dense captioning and visual grounding

    Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In ICCV,

    work page

Show all 85 references
  1. [9]

    4d spatio-temporal convnets: Minkowski convolutional neural networks

    Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 1, 2, 5, 7, 8, 12

    2019

  2. [10]

    MMDetection3D: Open- MMLab next-generation platform for general 3D object detection

    MMDetection3D Contributors. MMDetection3D: Open- MMLab next-generation platform for general 3D object detection. https://github.com/open- mmlab/ mmdetection3d, 2020. 12

    2020

  3. [11]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Hal- ber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 12

    2017

  4. [12]

    Autoen- coders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022

    Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jian- jian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoen- coders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 2, 7

    2022 arXiv

  5. [13]

    Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds

    Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. In NeurIPS, 2024. 2

    2024

  6. [14]

    Shape2scene: 3d scene representation learning through pre- training on shape data

    Tuo Feng, Wenguan Wang, Ruijie Quan, and Yi Yang. Shape2scene: 3d scene representation learning through pre- training on shape data. In ECCV, 2024. 2

    2024

  7. [15]

    Mamba3d: Enhancing local features for 3d point cloud anal- ysis via state space model.arXiv preprint arXiv:2404.14966,

    Xu Han, Yuan Tang, Zhaoxuan Wang, and Xianzhi Li. Mamba3d: Enhancing local features for 3d point cloud anal- ysis via state space model.arXiv preprint arXiv:2404.14966,

    arXiv

  8. [16]

    Momentum contrast for unsupervised visual rep- resentation learning

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual rep- resentation learning. In CVPR, 2020. 2

    2020

  9. [17]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2

    2022

  10. [18]

    Exploring data-efficient 3d scene understanding with contrastive scene contexts

    Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 5, 7, 8

    2021

  11. [19]

    Ponder: Point cloud pre-training via neural rendering

    Di Huang, Sida Peng, Tong He, Honghui Yang, Xiaowei Zhou, and Wanli Ouyang. Ponder: Point cloud pre-training via neural rendering. In ICCV, 2023. 2, 3, 7, 12

    2023

  12. [20]

    Spatio-temporal self-supervised representation learning for 3d point clouds

    Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In ICCV, 2021. 12

    2021

  13. [21]

    Pointgroup: Dual-set point grouping for 3d instance segmentation

    Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi- Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. In CVPR, 2020. 8

    2020

  14. [22]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3

    2023

  15. [23]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014. 5

    2014 arXiv

  16. [24]

    Oneformer3d: One transformer for unified point cloud segmentation

    Maxim Kolodiazhnyi, Anna V orontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In CVPR, 2024. 2

    2024

  17. [25]

    Stratified trans- former for 3d point cloud segmentation

    Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified trans- former for 3d point cloud segmentation. In CVPR, 2022. 2, 7

    2022

  18. [26]

    Masked discrimina- tion for self-supervised learning on point clouds

    Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimina- tion for self-supervised learning on point clouds. In ECCV,

  19. [27]

    Regress before construct: Regress autoen- coder for point cloud self-supervised learning

    Yang Liu, Chen Chen, Can Wang, Xulin King, and Mengyuan Liu. Regress before construct: Regress autoen- coder for point cloud self-supervised learning. In ACM MM,

  20. [28]

    Pointclustering: Unsupervised point cloud pre-training using transformation invariance in clustering

    Fuchen Long, Ting Yao, Zhaofan Qiu, Lusong Li, and Tao Mei. Pointclustering: Unsupervised point cloud pre-training using transformation invariance in clustering. In CVPR,

  21. [29]

    Decoupled weight decay regularization

    I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5

    2017 arXiv

  22. [30]

    Unified-io: A unified model for vision, language, and multi-modal tasks

    Jiasen Lu, Christopher Clark, Rowan Zellers, Roozbeh Mot- taghi, and Aniruddha Kembhavi. Unified-io: A unified model for vision, language, and multi-modal tasks. In ICLR,

  23. [31]

    Re- thinking network design and local geometry in point cloud: A simple residual mlp framework

    Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Re- thinking network design and local geometry in point cloud: A simple residual mlp framework. In ICLR, 2022. 2, 5, 6, 7, 8 9

    2022

  24. [32]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. Communications of the ACM, 2021. 3, 7

    2021

  25. [33]

    Masked autoencoders for point cloud self-supervised learning

    Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 1, 2, 5, 6, 7

    2022

  26. [34]

    Self-positioning point-based transformer for point cloud understanding

    Jinyoung Park, Sanghyeok Lee, Sihyeon Kim, Yunyang Xiong, and Hyunwoo J Kim. Self-positioning point-based transformer for point cloud understanding. In CVPR, 2023. 2

    2023

  27. [35]

    Oa-cnns: Omni- adaptive sparse cnns for 3d semantic segmentation

    Bohao Peng, Xiaoyang Wu, Li Jiang, Yukang Chen, Heng- shuang Zhao, Zhuotao Tian, and Jiaya Jia. Oa-cnns: Omni- adaptive sparse cnns for 3d semantic segmentation. InCVPR,

  28. [36]

    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. 1, 2, 7

    2017

  29. [37]

    Point- net++ deep hierarchical feature learning on point sets in a metric space

    Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point- net++ deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 1, 2, 7

    2017

  30. [38]

    Deep hough voting for 3d object detection in point clouds

    Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, 2019. 12

    2019

  31. [39]

    Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining

    Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023. 2

    2023 arXiv

  32. [40]

    Vpp: Efficient conditional 3d generation via voxel-point pro- gressive representation

    Zekun Qi, Muzhou Yu, Runpei Dong, and Kaisheng Ma. Vpp: Efficient conditional 3d generation via voxel-point pro- gressive representation. NeurIPS, 2024. 2

    2024

  33. [41]

    Shapellm: Universal 3d object understanding for embodied interaction

    Zekun Qi, Runpei Dong, Shaochen Zhang, Haoran Geng, Chunrui Han, Zheng Ge, Li Yi, and Kaisheng Ma. Shapellm: Universal 3d object understanding for embodied interaction. In ECCV, 2025. 2

    2025

  34. [42]

    Pointnext: Revisiting pointnet++ with improved training and scaling strategies

    Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 2, 6, 7

    2022

  35. [43]

    Randomrooms: Unsupervised pre- training from synthetic shapes and randomized layouts for 3d object detection

    Yongming Rao, Benlin Liu, Yi Wei, Jiwen Lu, Cho-Jui Hsieh, and Jie Zhou. Randomrooms: Unsupervised pre- training from synthetic shapes and randomized layouts for 3d object detection. In ICCV, 2021. 12

    2021

  36. [44]

    Bringing masked autoencoders explicit con- trastive properties for point cloud self-supervised learning

    Bin Ren, Guofeng Mei, Danda Pani Paudel, Weijie Wang, Yawei Li, Mengyuan Liu, Rita Cucchiara, Luc Van Gool, and Nicu Sebe. Bringing masked autoencoders explicit con- trastive properties for point cloud self-supervised learning. arXiv preprint arXiv:2407.05862, 2024. 2, 5, 6, 7

    2024 arXiv

  37. [45]

    High-resolution image syn- thesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image syn- thesis with latent diffusion models. In CVPR, 2022. 5

    2022

  38. [46]

    Language- grounded indoor 3d semantic segmentation in the wild

    David Rozenberszki, Or Litany, and Angela Dai. Language- grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 8

    2022

  39. [47]

    Language- grounded indoor 3d semantic segmentation in the wild

    David Rozenberszki, Or Litany, and Angela Dai. Language- grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 2, 6

    2022

  40. [48]

    Splatter image: Ultra-fast single-view 3d recon- struction

    Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d recon- struction. In CVPR, 2024. 3

    2024

  41. [49]

    Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, Franc ¸ois Goulette, and Leonidas J

    Hugues Thomas, Charles R. Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, Franc ¸ois Goulette, and Leonidas J. Guibas. Kpconv: Flexible and deformable convolution for point clouds. ICCV, 2019. 7

    2019

  42. [50]

    Kpconvx: Modernizing kernel point convolution with kernel attention

    Hugues Thomas, Yao-Hung Hubert Tsai, Timothy D Bar- foot, and Jian Zhang. Kpconvx: Modernizing kernel point convolution with kernel attention. In CVPR, 2024. 2

    2024

  43. [51]

    Revisiting point cloud classification: A new benchmark dataset and classifi- cation model on real-world data

    Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classifi- cation model on real-world data. In ICCV, 2019. 2, 6

    2019

  44. [52]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3, 5, 6, 7, 8

    2017

  45. [53]

    Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding

    Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bo- hao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, 2024. 2

    2024

  46. [54]

    Gpsformer: A global perception and local struc- ture fitting-based transformer for point cloud understanding

    Changshuo Wang, Meiqing Wu, Siew-Kei Lam, Xin Ning, Shangshu Yu, Ruiping Wang, Weijun Li, and Thambipillai Srikanthan. Gpsformer: A global perception and local struc- ture fitting-based transformer for point cloud understanding. arXiv preprint arXiv:2407.13519, 2024. 2

    2024 arXiv

  47. [55]

    Unsupervised point cloud pre-training via occlusion completion

    Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In ICCV, 2021. 6

    2021

  48. [56]

    Beyond first impressions: Integrating joint multi-modal cues for comprehensive 3d representation

    Haowei Wang, Jiji Tang, Jiayi Ji, Xiaoshuai Sun, Rongsheng Zhang, Yiwei Ma, Minda Zhao, Lincheng Li, Zeng Zhao, Tangjie Lv, et al. Beyond first impressions: Integrating joint multi-modal cues for comprehensive 3d representation. In ACM MM, pages 3403–3414, 2023. 2

    2023

  49. [57]

    Octformer: Octree-based transformers for 3d point clouds

    Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. TOG, 2023. 2, 7

    2023

  50. [58]

    Image as a foreign language: Beit pretraining for vision and vision- language tasks

    Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhil- iang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mo- hammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for vision and vision- language tasks. In CVPR, 2023. 1

    2023

  51. [59]

    Dynamic graph cnn for learning on point clouds

    Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. TOG, 2019. 7

    2019

  52. [60]

    Take-a-photo: 3d-to-2d generative pre-training of point cloud models

    Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models. In ICCV, 2023. 2, 3, 5, 6, 7

    2023

  53. [61]

    Point transformer v2: Grouped vector atten- tion and partition-based pooling

    Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Heng- shuang Zhao. Point transformer v2: Grouped vector atten- tion and partition-based pooling. In NeurIPS, 2022. 2, 7

    2022

  54. [62]

    Masked scene contrast: A scalable framework for unsuper- vised 3d representation learning

    Xiaoyang Wu, Xin Wen, Xihui Liu, and Hengshuang Zhao. Masked scene contrast: A scalable framework for unsuper- vised 3d representation learning. In CVPR, 2023. 2, 5, 7, 8

    2023

  55. [63]

    Point transformer v3: Simpler faster stronger

    Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xi- hui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang 10 Zhao. Point transformer v3: Simpler faster stronger. In CVPR, 2024. 2, 5, 7

    2024

  56. [64]

    Towards large- scale 3d representation learning with multi-dataset point prompt training

    Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large- scale 3d representation learning with multi-dataset point prompt training. In CVPR, 2024. 2, 7, 8

    2024

  57. [65]

    Pointcontrast: Unsupervised pre- training for 3d point cloud understanding

    Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pre- training for 3d point cloud understanding. In ECCV, 2020. 2, 5, 7, 8, 12

    2020

  58. [66]

    Disn: Deep implicit surface network for high-quality single-view 3d reconstruction

    Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In NeurIPS, 2019. 5

    2019

  59. [67]

    Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding

    Le Xue, Mingfei Gao, Chen Xing, Roberto Mart ´ın-Mart´ın, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In CVPR, 2023. 2

    2023

  60. [68]

    Ulip-2: Towards scal- able multimodal pre-training for 3d understanding

    Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Jun- nan Li, Roberto Mart´ın-Mart´ın, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scal- able multimodal pre-training for 3d understanding. InCVPR,

  61. [69]

    Point cloud pre- training with natural 3d structures

    Ryosuke Yamada, Hirokatsu Kataoka, Naoya Chiba, Yukiyasu Domae, and Tetsuya Ogata. Point cloud pre- training with natural 3d structures. In CVPR, 2022. 12

    2022

  62. [70]

    Implicit autoencoder for point-cloud self-supervised representation learning

    Siming Yan, Zhenpei Yang, Haoxiang Li, Chen Song, Li Guan, Hao Kang, Gang Hua, and Qixing Huang. Implicit autoencoder for point-cloud self-supervised representation learning. In ICCV, 2023. 12

    2023

  63. [71]

    Swin3d: A pretrained transformer backbone for 3d indoor scene understanding

    Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv preprint arXiv:2304.06906 ,

    arXiv

  64. [72]

    A scalable active framework for region annotation in 3d shape collections

    Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Shef- fer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ToG, 2016. 2, 6

    2016

  65. [73]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling

    Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 5, 6, 7

    2022

  66. [74]

    Towards compact 3d representations via point feature enhancement masked au- toencoders

    Yaohua Zha, Huizhen Ji, Jinmin Li, Rongsheng Li, Tao Dai, Bin Chen, Zhi Wang, and Shu-Tao Xia. Towards compact 3d representations via point feature enhancement masked au- toencoders. In AAAI, 2024. 2

    2024

  67. [75]

    Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training

    Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. NeurIPS, 2022

    2022

  68. [76]

    Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders

    Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hong- sheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In CVPR,

  69. [77]

    Point could mamba: Point cloud learning via state space model

    Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024. 2, 5, 6, 8

    2024 arXiv

  70. [78]

    Pcp- mae: Learning to predict centers for point masked autoen- coders

    Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcp- mae: Learning to predict centers for point masked autoen- coders. arXiv preprint arXiv:2408.08753, 2024. 2, 7

    2024 arXiv

  71. [79]

    Meta- transformer: A unified framework for multimodal learning

    Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta- transformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 1

    2023 arXiv

  72. [80]

    Self-supervised pretraining of 3d features on any point-cloud

    Zaiwei Zhang, Rohit Girdhar, Armand Joulin, and Ishan Misra. Self-supervised pretraining of 3d features on any point-cloud. In ICCV, 2021. 12

    2021

  73. [81]

    Point transformer

    Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 2, 7

    2021

  74. [82]

    Point cloud pre-training with diffusion models

    Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In CVPR,

  75. [83]

    Uni3d: Exploring unified 3d representation at scale

    Junsheng Zhou, Jinsheng Wang, Baorui Ma, Yu-Shen Liu, Tiejun Huang, and Xinlong Wang. Uni3d: Exploring unified 3d representation at scale. arXiv preprint arXiv:2310.06773,

    arXiv

  76. [84]

    Ponderv2: Pave the way for 3d foundataion model with a universal pre-training paradigm

    Haoyi Zhu, Honghui Yang, Xiaoyang Wu, Di Huang, Sha Zhang, Xianglong He, Tong He, Hengshuang Zhao, Chun- hua Shen, Yu Qiao, et al. Ponderv2: Pave the way for 3d foundataion model with a universal pre-training paradigm. arXiv preprint arXiv:2310.08586, 2023. 2, 3, 7, 8

    arXiv

  77. [85]

    Uni-perceiver: Pre- training unified architecture for generic perception for zero- shot and few-shot tasks

    Xizhou Zhu, Jinguo Zhu, Hao Li, Xiaoshi Wu, Hongsheng Li, Xiaohua Wang, and Jifeng Dai. Uni-perceiver: Pre- training unified architecture for generic perception for zero- shot and few-shot tasks. In CVPR, 2022. 1 11 UniPre3D: Unified Pre-training of 3D Point Cloud Models with ...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.