Pith. sign in

REVIEW 4 major objections 6 minor 50 references

TriCLIP-3D: A Unified Parameter-Efficient Framework for Tri-Modal 3D Visual Grounding based on CLIP

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read One frozen CLIP model with lightweight adapters encodes point clouds, images, and text for 3D detection and grounding, improving on EmbodiedScan by 6.52 AP25 in detection and 6.25 AP25 in grounding with 58% fewer trainable parameters.

desk verdict A useful incremental idea with an unproven headline number because the baseline's test-time protocol is never specified. read the letter →

arxiv 2507.14904 v2 pith:DNPLT7NJ submitted 2025-07-20 cs.CV cs.AI

Fan Li , Zanyi Wang , Zeyi Huang , Guang Dai , Jingdong Wang , Mengmeng Wang This is my paper
classification cs.CVcs.AI
keywords 3Dvisualgroundingobjectdetectionmulti-modalfusionCLIPpointcloudparameter-efficientfine-tuningresidualadapterEmbodiedScan
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

TriCLIP-3D proposes to replace the usual stack of separate 2D, 3D, and text encoders with one frozen CLIP model: the same CLIP vision transformer encodes both multi-view images and tokenized point clouds, while the CLIP text encoder handles the instruction. The motivation is that separate encoders create misaligned feature spaces and parameter waste, and prior CLIP-for-3D methods still kept dedicated 3D backbones. On EmbodiedScan, the paper reports a gain of 6.52 AP25 on 3D detection, 6.25 AP25 on 3D visual grounding, and roughly 58% fewer trainable parameters than the EmbodiedScan baseline. A geometric recovery and fusion module (GARF) is the piece that keeps 3D structure from being lost when CLIP tokens are fused with images. If the result holds, expensive dedicated 3D encoders may not be needed for this class of embodied perception tasks.

What carries the argument

The load-bearing object is the unified frozen CLIP encoder with shared CLIP-ViT weights for images and point clouds, plus per-modality residual adapters. Point tokenization follows EPCL: farthest point sampling, K-nearest-neighbor grouping into M patches, and an MLP embedding into D-dimensional tokens, so point patches enter the same sequence interface as image patches. The GARF module is the second piece: it re-materializes sparse 3D tensors and dense 2D feature maps from CLIP tokens, projects 3D features onto images via camera intrinsics and extrinsics, and fuses them with the Adaptive Point-Image Fusion (APIF) module, a pooling-and-shared-MLP gating mechanism. Together these pieces let the same pretrained network serve text, image, and point-cloud modalities and feed a multi-modal decoder for grounding.

What would settle it

A controlled comparison on EmbodiedScan where the only change is replacing the frozen CLIP-ViT point-cloud branch with a randomly initialized ViT of the same shape, keeping adapters, GARF, and decoder fixed; if detection and grounding AP25 do not drop sharply, the claim that CLIP's pretrained knowledge is what makes the unified encoder work would be falsified.

Watch

Extended reading notes

Core claim

The central claim is that a 2D pretrained CLIP model can be extended into a unified tri-modal encoder for 3D visual grounding: point clouds are tokenized into patches using farthest point sampling, K-nearest-neighbor grouping, and an MLP embedding (following EPCL), and those tokens go through the same frozen CLIP vision transformer that processes multi-view RGB images, while text goes through the frozen CLIP text encoder. Per-modality residual adapters inserted in odd transformer layers let each modality fine-tune without unfreezing CLIP. Because raw CLIP tokens lose explicit 3D geometry, the Geometric-Aware 2D-3D Feature Recovery and Fusion (GARF) module recovers multi-scale 3D sparse tensors from point features and multi-scale 2D feature maps from image features, projects 3D points into image space using camera parameters, and adaptively fuses them with the APIF module. The fused features feed a 3D detection head or a multi-modal decoder that also crosses attention with image and text features. On EmbodiedScan this yields 23.37 AP25 detection, a gain of 6.52 over the baseline, plus grounding gains of 6.25 AP25 on the mini split and 6.36 AP25 on the full split, with total trainable parameters reduced to about 42% of the baseline.

Load-bearing premise

The load-bearing premise is that a point cloud, once cut into local patches, behaves enough like a set of image patches that a frozen 2D CLIP vision transformer can extract features strong enough to support 3D detection and grounding.

Editorial extensions

If this is right

  • The feature-extraction stage no longer needs a dedicated 3D backbone: point clouds enter the same frozen CLIP-ViT as images, and the reported gains come on top of removing that network.
  • Trainable parameters fall to about 42% of the EmbodiedScan baseline (96.72M vs 229.55M), so the improvement is not bought with extra capacity.
  • GARF is the main source of grounding accuracy in the ablation: removing it drops AP25 from 39.84 to 29.45 on the mini split, indicating that geometric recovery is what makes naive CLIP-token fusion viable.
  • The largest detection gains land on tail categories (+8.53 AP25), and grounding gains are largest on view-dependent examples, consistent with CLIP's pretrained multimodal knowledge transferring to rare and language-dependent objects.

Reading between the lines

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

  • We infer that the same frozen-CLIP-with-adapters recipe could transfer to other 3D perception tasks such as instance segmentation, dense captioning, or open-vocabulary detection, since the paper only evaluates detection and grounding.
  • We infer that the reported comparison may be sensitive to input protocol, including 20 training views, 50 test views, and 100k sampled points; a controlled rerun of EmbodiedScan under exactly this protocol would clarify how much of the gain is architectural.
  • We infer a test to separate CLIP's semantic priors from geometric reasoning: corrupt point coordinates by rotation, jitter, or patch shuffling while keeping CLIP frozen; if AP25 stays high, the method is relying on semantics rather than 3D geometry.
  • We infer that if this recipe holds, future embodied systems could share one vision-language backbone across modalities instead of maintaining per-modality encoders, simplifying deployment on robots with limited compute.
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

4 major / 6 minor

Summary. The paper proposes TriCLIP-3D, a framework for 3D object detection and 3D visual grounding on the EmbodiedScan benchmark. The core idea is to use a single frozen CLIP model, with lightweight residual adapters, as the encoder for multi-view images, point clouds, and text, thereby removing a dedicated 3D backbone. A Geometric-Aware 2D-3D Feature Recovery and Fusion (GARF) module recovers 3D and 2D structure from CLIP features and fuses them, and a multi-modal decoder incorporates image features to improve grounding. The paper reports consistent AP improvements over EmbodiedScan on both tasks, a reduction in trainable parameters by about 58%, and ablations of the GARF module and the decoder.

Significance. If the reported gains are robust, the paper would offer a useful demonstration that a frozen 2D CLIP encoder can support point-cloud-based 3D detection and grounding, and that adapters plus geometric fusion can compensate for the loss of a dedicated 3D backbone. The parameter-efficiency claim is attractive for embodied perception. The paper also provides a reasonably clean presentation and ablates its two named modules. However, the central quantitative claims currently rest on an incompletely specified comparison, and several internal inconsistencies and unexplained per-dataset results prevent me from taking the headline numbers at face value.

major comments (4)
  1. [§1, §3.2(2), Eq. (5)] The central claim that the method eliminates the need for a separate 3D network is contradicted by the method description. In §3.2(2), the point cloud is first processed by 3D Minkowski Convolution to produce S1 = SparseTensor(F1,U1), and Eq. (5) again applies 3D Minkowski convolution to the concatenated sparse tensor to generate multi-scale sparse features. These are 3D networks performing geometric feature extraction before and after the CLIP ViT. Please either remove the 'no 3D network' claim, justify why these Minkowski layers are not counted as a 3D backbone, or report their parameter cost separately in Table 3.
  2. [§4.2, Tables 1–2] The headline comparison with EmbodiedScan is not matched-protocol. Section 4.2 specifies TriCLIP-3D's inputs (20 training views, 50 test views, 224x224 resolution, 100k points) but gives no equivalent protocol for the EmbodiedScan baseline. If the baseline was evaluated with fewer views or fewer points, the reported +6.52 AP25 and +6.25 AP25 gains could reflect the additional input information rather than the proposed unified encoder. The baseline's input protocol must be reported, or the baseline must be re-run under the same protocol, before the central quantitative claim is established.
  3. [Table 4] The per-dataset ablation contradicts the claim of consistent generalization. For 3D detection, TriCLIP-3D improves over EmbodiedScan by +19.56 AP25 on 3RScan but only +0.08 AP25 on Matterport3D, while the grounding gains in Table 5 are +4.49 and +2.13 respectively. This pattern is not explained anywhere in the paper and is more consistent with a weak baseline or protocol mismatch on 3RScan than with a generalizable architectural benefit. The authors should analyze this discrepancy, provide matched-protocol per-dataset results, and explain why the 3RScan gain is so much larger than the gains on the other two datasets.
  4. [§4.3, Table 1] The reported AP gains are accompanied by substantial recall decreases, which weakens the claim of an unequivocal improvement. In Table 1, the Large-Vocabulary AR25 drops from 51.07 to 47.70, head AR25 drops from 67.51 to 62.80, and common AR25 drops from 50.46 to 46.17. The paper should report precision-recall behavior or otherwise justify why the AP25 improvement is preferable despite lower recall, especially since the abstract describes the result as a general accuracy improvement.
minor comments (6)
  1. [Abstract, §4.3] The phrase '6.52% improvement' should be '6.52 percentage-point improvement in AP25'; the same applies to the AP50 and grounding numbers. AP is already a percentage metric, so reporting it as a '% improvement' is misleading.
  2. [Table 5] The table title reads 'Ablation Study on 3D visual detection Performance Across Various Datasets' but the task is 3D visual grounding; the title should be corrected.
  3. [§3.2(2), §4.2] The text in §3.2(2) refers to a 'FPS K value', but K is the number of neighbors in the KNN grouping step, not in FPS. In §4.2, the K value is 16 and the number of point-sample groups is 512; please clarify which parameter is which.
  4. [§3.3, Eq. (6)] The notation for the reshaped image features f_i_r is imprecise: the relationship between the CLIP feature length L, the number of views, and the feature map dimensions C_i x H_i x W_i is not defined. Please provide the exact reshape mapping.
  5. [§3.4, Eq. (12), §4.2] The loss weights alpha, beta, and gamma are said to be 'typically set to 1.0', but the exact values used for the reported results are not specified. Please report the exact weights for both tasks, and ideally a sensitivity analysis.
  6. [References] Several references are incomplete or include raw arXiv/semanticscholar URLs (e.g., [24], [34], [35]); these should be formatted consistently in the bibliography style.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are validated against the external EmbodiedScan benchmark and controlled ablations; no equation reduces to its own inputs and no self-citation is load-bearing.

full rationale

The paper is an empirical method study with no mathematical derivation chain to reduce. Every central claim — unified CLIP encoding of point clouds, GARF fusion, multi-modal decoder, parameter reduction, and the headline AP gains — is validated against the external EmbodiedScan benchmark under standard metrics (AP25/AP50/AR), and the ablations in Tables 4-7 are internally consistent comparisons of the proposed components against their own ablated versions. The load-bearing premise that point clouds can be tokenized into patches and fed into a frozen CLIP ViT (Section 3.2(2)) is explicitly adopted from EPCL [19], an independent external prior work by different authors, and is empirically tested rather than assumed by definition; the paper's own contributions (GARF, residual adapters, decoder) are evaluated with controlled ablations showing the deltas. No equation in the paper defines a target metric in terms of the model's own outputs, no fitted parameter is renamed as a prediction, and no self-citation or uniqueness theorem is invoked to force a choice, so the derivation is self-contained. The skeptical concern — that the EmbodiedScan baseline's test-time protocol (view count, resolution, point sampling) is not specified while TriCLIP-3D's is (Section 4.2) — is a legitimate experimental-fairness and correctness risk, not a circularity, since the reported gains are not forced by construction; indeed Table 1's simultaneous AR25 drop (51.07 to 47.70) indicates a measured precision/recall tradeoff rather than a tautological improvement. The paper contains no self-citations at all, and its reliance on external prior work (EPCL, SENet, EmbodiedScan) is transparent and testable, so the honest finding is no circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The paper does not introduce physical entities. The free parameters are standard design choices, and the axioms are domain assumptions about CLIP transfer, point tokenization, and benchmark comparability.

free parameters (5)
  • Loss weights alpha, beta, gamma = 1.0 each
    Set by hand in Section 3.4 (Eq. 12) for the combined loss; not tuned.
  • Number of sampled points per scene = 100,000
    Chosen in Section 4.2; affects input to the point tokenizer.
  • Number of input views = 20 training, 50 testing
    Chosen in Section 4.2; the baseline protocol is not specified, so this may affect comparability.
  • FPS K value and number of point groups = K=16, groups=512
    Chosen in Section 4.2 for point patch generation; affects the token sequence length fed to CLIP.
  • Hidden size = 768
    Chosen in Section 4.2 to match CLIP ViT-B/16 feature dimension.
assumptions (4)
  • domain assumption Point clouds can be tokenized into patches and processed by a 2D CLIP ViT (FPS+KNN+MLP tokenization)
    Inherited from EPCL [19]; central to the unified encoder design in Section 3.2(2).
  • domain assumption CLIP pretrained representations transfer to 3D detection and grounding tasks after adapter fine-tuning
    Core premise of the paper, stated in the introduction and Section 3.2; not proven beyond the reported benchmark results.
  • domain assumption Camera intrinsics and extrinsics are known for 3D-to-2D projection in GARF
    Standard in EmbodiedScan; used in Section 3.3(3).
  • domain assumption The EmbodiedScan baseline and evaluation protocol are reproducible and comparable
    The paper compares against published EmbodiedScan results but does not specify the baseline input protocol (number of views, resolution, sampling), so comparability is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TriCLIP-3D: A Unified Parameter-Efficient Framework for Tri-Modal 3D Visual Grounding based on CLIP." pith.science (2026). https://pith.science/paper/DNPLT7NJ

@misc{pith2026250714904,
  author       = {Pith},
  title        = {Pith review of: TriCLIP-3D: A Unified Parameter-Efficient Framework for Tri-Modal 3D Visual Grounding based on CLIP},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DNPLT7NJ}},
  note         = {Machine review of arXiv:2507.14904}
}
read the original abstract

3D visual grounding allows an embodied agent to understand visual information in real-world 3D environments based on human instructions, which is crucial for embodied intelligence. Existing 3D visual grounding methods typically rely on separate encoders for different modalities (e.g., RGB images, text, and 3D point clouds), resulting in large and complex models that are inefficient to train. While some approaches use pre-trained 2D multi-modal models like CLIP for 3D tasks, they still struggle with aligning point cloud data to 2D encoders. As a result, these methods continue to depend on 3D encoders for feature extraction, further increasing model complexity and training inefficiency. In this paper, we propose a unified 2D pre-trained multi-modal network to process all three modalities (RGB images, text, and point clouds), significantly simplifying the architecture. By leveraging a 2D CLIP bi-modal model with adapter-based fine-tuning, this framework effectively adapts to the tri-modal setting, improving both adaptability and performance across modalities. Our Geometric-Aware 2D-3D Feature Recovery and Fusion (GARF) module is designed to fuse geometric multi-scale features from point clouds and images. We then integrate textual features for final modality fusion and introduce a multi-modal decoder to facilitate deep cross-modal understanding. Together, our method achieves unified feature extraction and fusion across the three modalities, enabling an end-to-end 3D visual grounding model. Compared to the baseline, our method reduces the number of trainable parameters by approximately 58\%, while achieving a 6.52\% improvement in the 3D detection task and a 6.25\% improvement in the 3D visual grounding task.

Figures

Figures reproduced from arXiv: 2507.14904 by the authors.

Figure 1
Figure 1. Comparison of different encoder architectures: (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our framework. The proposed network architecture integrates multiple modalities, including RGB [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) The Geometric-Aware 2D-3D Feature Recovery [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The left part of the figure illustrates the 3D visual [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of 3D detection task [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of 3D visual grounding task [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 32 canonical work pages

  1. [1]

    Ahmed Abdelreheem, Ujjwal Upadhyay, Ivan Skorokhodov, Rawan Al Yahya, Jun Chen, and Mohamed Elhoseiny. 2022. 3dreftransformer: Fine-grained object 8 TriCLIP-3D: A Unified Parameter-Efficient Framework for Tri-Modal 3D Visual Grounding based on CLIP identification in real-world scenes using natural language. In Proceedings of the IEEE/CVF winter conference...

    work page 2022

  2. [2]

    Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. 2020. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In Computer Vision–ECCV 2020: 16th Euro- pean Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part I 16 . Springer, 422–440

    work page 2020

  3. [3]

    Eslam Bakr, Yasmeen Alsaedy, and Mohamed Elhoseiny. 2022. Look around and refer: 2d synthetic semantics knowledge distillation for 3d visual grounding. Advances in neural information processing systems 35 (2022), 37146–37158

    work page 2022

  4. [4]

    Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 2022. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 16464–16473

    2022

  5. [5]

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexan- der Kirillov, and Sergey Zagoruyko. 2020. End-to-end object detection with transformers. In European conference on computer vision . Springer, 213–229

    2020

  6. [6]

    Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. 2017. Matterport3D: Learning from RGB-D Data in Indoor Environments. International Conference on 3D Vision (3DV) (2017)

    2017

  7. [7]

    Chun-Peng Chang, Shaoxiang Wang, Alain Pagani, and Didier Stricker. 2024. MiKASA: Multi-key-anchor & scene-aware transformer for 3d visual ground- ing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 14131–14140

    work page 2024

  8. [8]

    Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. 2020. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision. Springer, 202–221

    2020

Show all 50 references
  1. [9]

    Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang

  2. [10]

    Christopher Choy, JunYoung Gwak, and Silvio Savarese. 2019. 4D Spatio- Temporal ConvNets: Minkowski Convolutional Neural Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition . 3075–3084

    2019

  3. [11]

    Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner

    Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. 2017. ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE

    2017

  4. [12]

    Mingtao Feng, Zhen Li, Qi Li, Liang Zhang, XiangDong Zhang, Guangming Zhu, Hui Zhang, Yaonan Wang, and Ajmal Mian. 2021. Free-form description guided 3d visual graph network for object grounding in point cloud. In Proceedings of the IEEE/CVF international conference on compute...

    2021

  5. [13]

    Shuvozit Ghose, Manyi Li, Yiming Qian, and Yang Wang. 2025. CLIP-Based Point Cloud Classification via Point Cloud to Image Translation. In International Conference on Pattern Recognition . Springer, 173–186

    2025

  6. [14]

    Deepti Hegde, Jeya Maria Jose Valanarasu, and Vishal Patel. 2023. Clip goes 3d: Leveraging prompt tuning for language grounded 3d recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2028–2038

    2023

  7. [15]

    Jie Hu, Li Shen, and Gang Sun. 2018. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition . 7132–7141

    2018

  8. [16]

    Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. 2022. Multi-view trans- former for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 15524–15533

    2022

  9. [17]

    Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. 2023. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. InProceedings of the IEEE/CVF International Conference on Computer Vision . 22157–22167

    2023

  10. [18]

    Wencan Huang, Daizong Liu, and Wei Hu. 2023. Dense object grounding in 3d scenes. In Proceedings of the 31st ACM International Conference on Multimedia . 5017–5026

    2023

  11. [19]

    Xiaoshui Huang, Zhou Huang, Sheng Li, Wentao Qu, Tong He, Yuenan Hou, Yifan Zuo, and Wanli Ouyang. 2024. EPCL: Frozen CLIP Transformer is An Efficient Point Cloud Encoder. In Proceedings of the AAAI Conference on Artificial Intelligence

    2024

  12. [20]

    Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki

  13. [21]

    Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. 2020. Pointgroup: Dual-set point grouping for 3d instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and Pattern recognition . 4867–4876

    2020

  14. [22]

    Zhenxiang Lin, Xidong Peng, Peishan Cong, Ge Zheng, Yujin Sun, Yuenan Hou, Xinge Zhu, Sibei Yang, and Yuexin Ma. 2024. Wildrefer: 3d object localization in large-scale dynamic scenes with multi-modal visual data and natural language. In European Conference on Computer Vision ....

    2024

  15. [23]

    Yang Liu, Daizong Liu, and Wei Hu. 2025. Joint Top-Down and Bottom-Up Frameworks for 3D Visual Grounding. In International Conference on Pattern Recognition. Springer, 249–264

    2025

  16. [24]

    Ilya Loshchilov and Frank Hutter. 2017. Fixing Weight Decay Regularization in Adam. ArXiv abs/1711.05101 (2017). https://api.semanticscholar.org/CorpusID: 3312944

    2017 arXiv

  17. [25]

    Taiki Miyanishi, Daichi Azuma, Shuhei Kurita, and Motoaki Kawanabe. 2024. Cross3dvg: Cross-dataset 3d visual grounding on different rgb-d scans. In 2024 International Conference on 3D Vision (3DV) . IEEE, 717–727

    2024

  18. [26]

    Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. 2022. Slip: Self-supervision meets language-image pre-training. In European conference on computer vision. Springer, 529–544

    2022

  19. [27]

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing system...

    2019

  20. [28]

    Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. 2019. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision . 9277–9286

    2019

  21. [29]

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. 2017. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems 30 (2017)

    2017

  22. [30]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

  23. [31]

    Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. 2022. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning. PMLR, 1046–1056

    2022

  24. [32]

    T-YLPG Ross and GKHP Dollár. 2017. Focal loss for dense object detection. In proceedings of the IEEE conference on computer vision and pattern recognition . 2980–2988

    2017

  25. [33]

    Danila Rukhovich, Anna Vorontsova, and Anton Konushin. 2022. Fcaf3d: Fully convolutional anchor-free 3d object detection. In European Conference on Com- puter Vision. Springer, 477–493

    2022

  26. [34]

    Rukhovich, Anna Vorontsova, and Anton Konushin

    Danila D. Rukhovich, Anna Vorontsova, and Anton Konushin. 2021. ImVoxelNet: Image to Voxels Projection for Monocular and Multi-View General-Purpose 3D Object Detection. 2022 IEEE/CVF Winter Conference on Applications of Computer Vision (W ACV)(2021), 1265–1274. https://api.sem...

    2021

  27. [35]

    Lang, Bassam Helou, and Oscar Beijbom

    Sourabh Vora, Alex H. Lang, Bassam Helou, and Oscar Beijbom. 2019. Point- Painting: Sequential Fusion for 3D Object Detection. 2020 IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) (2019), 4603–4611. https://api.semanticscholar.org/CorpusID:208248084

    2019

  28. [36]

    Johanna Wald, Armen Avetisyan, Nassir Navab, Federico Tombari, and Matthias Niessner. 2019. RIO: 3D Object Instance Re-Localization in Changing Indoor Environments. In Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    2019

  29. [37]

    Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. 2024. EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI. In IEEE Conference on ...

    2024

  30. [38]

    Changli Wu, Yiwei Ma, Qi Chen, Haowei Wang, Gen Luo, Jiayi Ji, and Xiaoshuai Sun. 2024. 3d-stmn: Dependency-driven superpoint-text matching network for end-to-end 3d referring expression segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 5940–5948

    2024

  31. [39]

    Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. 2023. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 19231–19242

    2023

  32. [40]

    Le Xue, Mingfei Gao, Chen Xing, Roberto Martín-Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. 2023. Ulip: Learning a unified representation of language, images, and point clouds for 3d understand- ing. In Proceedings of the IEEE/CVF confere...

    2023

  33. [41]

    Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martín- Martín, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. 2024. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer...

    2024

  34. [42]

    Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. 2021. Sat: 2d semantics assisted training for 3d visual grounding. InProceedings of the IEEE/CVF International Conference on Computer Vision . 1856–1866

    2021

  35. [43]

    Zhenfei Yin, Jiong Wang, Jianjian Cao, Zhelun Shi, Dingning Liu, Mukai Li, Xiaoshui Huang, Zhiyong Wang, Lu Sheng, Lei Bai, et al. 2023. Lamm: Language- assisted multi-modal instruction-tuning dataset, framework, and benchmark. Advances in Neural Information Processing Systems...

    2023

  36. [44]

    Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. 2024. Visual programming for zero-shot open-vocabulary 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20623–20633

    2024

  37. [45]

    Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. 2022. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8552–8562

    2022

  38. [46]

    Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 2021. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 2928–2937

    2021

  39. [47]

    Chenming Zhu, Wenwei Zhang, Tai Wang, Xihui Liu, and Kai Chen. 2023. Ob- ject2scene: Putting objects in context for open-vocabulary 3d detection. arXiv preprint arXiv:2309.09456 (2023). 10

    2023 arXiv

  40. [2021]

    In International conference on machine learning

    Learning transferable visual models from natural language supervision. In International conference on machine learning . PmLR, 8748–8763

  41. [2022]

    In European Conference on Computer Vision

    Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision . Springer, 417–433

  42. [2023]

    In Proceedings of the IEEE/CVF international conference on computer vision

    Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF international conference on computer vision . 18109– 18119

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.