Pith. sign in

REVIEW 6 major objections 5 minor 39 references

Unified Representation Space for 3D Visual Grounding

T0 review · 6 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 UniSpace-3D grounds 3D objects by mapping point clouds and text into one shared CLIP space, beating baselines by at least 2.24 points on ScanRefer, Nr3D, and Sr3D.

desk verdict A credible, incremental 3D visual grounding paper that combines known modules and reports consistent gains over a reproduced baseline, but the central 'unified space' claim is under-tested because the CLIP tokenization is underspecified and the ablations don't isolate the pretrained weights. read the letter →

arxiv 2506.14238 v1 pith:YFPNZ7ES submitted 2025-06-17 cs.CV

Yinuo Zheng , Lipeng Gu , Honghua Chen , Liangliang Nan , Mingqiang Wei This is my paper · ORCID
classification cs.CV
keywords 3DvisualgroundingunifiedrepresentationspaceCLIPcontrastivelearninglanguage-guidedqueryselectionpointcloudsmultimodalalignmentScanRefer
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

UniSpace-3D claims that the recurring failure mode of 3D visual grounding—correctly classifying an object but placing its box in the wrong spot—can be reduced by forcing the visual features of a point cloud and the textual features of a description into a single shared representation space. The paper builds this space with a frozen CLIP model that accepts point cloud patches tokenized as 2D manifolds, then adds contrastive learning and language-guided query selection. On the ScanRefer, Nr3D, and Sr3D benchmarks, the resulting model outperforms its reproduced baseline EDA by margins of roughly 2 to 5 accuracy points, with the biggest gains on hard, multi-instance scenes. A sympathetic reader would take the paper's contribution to be a recipe: reuse an image-text pretrained space rather than training separate encoders, and let language tokens steer which seed points become object candidates.

What carries the argument

The central object is the unified representation (UR) space, implemented by a frozen CLIP image transformer that treats point cloud patches as 2D manifolds and outputs task-position dual-aware visual embeddings in the same space as CLIP text embeddings. Around it, the paper builds two mechanisms: the multi-modal contrastive learning module, whose losses pull matched visual-text pairs together and push mismatched scenes and LLM-generated negative sentences apart, and the language-guided query selection module, which uses the maximum similarity between each visual query and all language queries to select the top $N_q$ seed points that initialize object candidate queries for the cross-modal decoder.

What would settle it

Train UniSpace-3D with the frozen CLIP image transformer replaced by a randomly initialized transformer of the same architecture, keeping all other components and losses identical. If ScanRefer Acc@0.5 stays near 43.95%, the unified representation space is not carrying the claimed signal; if it falls back toward the 41.71% baseline, the frozen CLIP space is doing the work.

Watch

Extended reading notes

Core claim

The paper's central claim is that aligning 3D visual features and text in the representation space of a pre-trained CLIP model removes a modality gap that prior one-stage methods leave in place, and that this alignment, not any new geometric reasoning, is what lifts grounding accuracy. Concretely, UniSpace-3D feeds local point cloud patches through PointNet into a frozen CLIP image transformer, feeds the text through the CLIP text encoder, and trains three auxiliary objectives—a positive contrastive loss pulling matched visual and textual embeddings together, negative contrastive losses that push embeddings away from wrong scenes and wrong sentences generated by GPT-3, and a language-guided query selection step that picks the top 256 visual queries by their maximum similarity to language queries. The reported outcome is 56.04% Acc@0.25 and 43.95% Acc@0.5 on ScanRefer overall, 57.2% on Nr3D, and 69.8% on Sr3D, exceeding the reproduced EDA baseline by at least 2.24 points. The authors attribute the improvement to the unified representation space itself and show through ablations that each added component contributes positively.

Load-bearing premise

The load-bearing premise is that a frozen image-text model trained on RGB photos can read point-cloud patches, tokenized as flat 2D surfaces, well enough to carry the spatial and semantic information grounding needs; the paper never tests that premise alone, only inside the full pipeline.

Editorial extensions

If this is right

  • If the unified space is the source of the gains, the same frozen-CLIP recipe should transfer to other 3D vision-language tasks such as dense captioning and 3D question answering, since they share the same vision-text alignment bottleneck.
  • Replacing the GPT-3-generated negative sentences with negatives from a smaller open model should still work as long as the negatives fit the scene context, because the loss only needs plausible distractors.
  • Language-guided query selection should make grounding degrade more gracefully when text is verbose or contains ordinal and spatial phrases, since language tokens directly influence which seed points are decoded.
  • Because the CLIP image transformer stays frozen, the added compute over a one-stage baseline is concentrated in the contrastive losses and query selection, making the accuracy gains comparatively inexpensive to reproduce.

Reading between the lines

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

  • The paper never isolates the frozen-CLIP assumption; one can test it by swapping the frozen CLIP image transformer for a randomly initialized transformer of identical shape while keeping every other component fixed, and checking whether accuracy collapses.
  • A stronger test of the unified space would be a cross-modal retrieval experiment: given a description, does the top visual embedding retrieved from the scene match the ground-truth object? That would measure the modality gap directly instead of through downstream box accuracy.
  • The loss weights $\alpha=0.5$, $\beta=0.3$, $\gamma=0.1$ are tuned on ScanRefer validation; transferring them to Nr3D/Sr3D assumes the three objectives stay in balance across datasets with very different sentence styles, which the paper does not ablate.
  • If the approach generalizes, its natural next target is open-vocabulary grounding, because the shared CLIP space already carries open-set semantics that the downstream decoder could be trained to read.
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

6 major / 5 minor

Summary. The paper proposes UniSpace-3D, a 3D visual grounding method whose core idea is to map point-cloud and text features into a shared representation space built on a frozen CLIP model. The method adds three components to an EDA-style baseline: a unified representation encoder (URE) that feeds point-cloud tokens through a frozen CLIP image transformer, a multi-modal contrastive learning module (MMCL) with positive and negative losses, and a language-guided query selection module (LGQS) that initializes decoder queries from text-conditioned similarity scores. The authors report results on ScanRefer and Nr3D/Sr3D, claiming improvements over EDA and prior state-of-the-art methods by at least 2.24 percentage points, and they provide incremental ablations plus qualitative examples.

Significance. If the claims are substantiated, the paper would make a useful empirical contribution by showing how a frozen CLIP encoder can be adapted to 3D point-cloud grounding, and by demonstrating that contrastive alignment and language-guided query selection give small but consistent gains over a reproduced EDA baseline. The inclusion of a reproduced baseline, the incremental ablation structure, and the qualitative failure cases are strengths. However, the central mechanism is not yet isolated: the URE ablation conflates the pretrained CLIP transformer with several new trainable components, the tokenization into CLIP-compatible tokens is underspecified, no error bars or significance tests are reported, and several reported numbers are inconsistent. The current evidence is not sufficient to establish the paper's central claim that the unified representation space, rather than added capacity or training objectives, drives the improvements.

major comments (6)
  1. [Sec. III-A3 and Fig. 2] The URE tokenization is underspecified. The text says PointNet maps local point cloud patches, 'viewed as 2D manifolds', into the vision token space and that the resulting tokens, together with task and position tokens, are fed into a frozen CLIP image transformer. It does not specify how the point cloud is partitioned into patches, how the patches are ordered, how many tokens are produced, or how CLIP's fixed grid positional embeddings are adapted to unordered point-cloud tokens. Without these details, the core architecture cannot be reproduced and the 'unified representation space' claim cannot be evaluated.
  2. [Tab. IV, rows (a)-(b)] The ablation labeled URE is not an ablation of the frozen CLIP encoder. Row (b) adds URE, but this also introduces the trainable MLPs, task tokenizer, position tokenizer, and any token-order or positional adaptations required to feed point-cloud tokens into CLIP. The reported gain from row (a) to row (b) therefore conflates the pretrained CLIP transformer with these new trainable components. A control experiment with a randomly initialized frozen transformer of the same capacity is needed to establish that the pretrained CLIP alignment, rather than additional capacity or the contrastive objective, is load-bearing.
  3. [Sec. IV-D and Tabs. I-IV] All reported numbers are single runs, the loss weights in Tab. III are selected on the same ScanRefer validation set that is reported, and no error bars or significance tests are provided. Several margins are small (e.g., +2.24 overall Acc@0.5 in Tab. I and component increments around +0.2 to +0.5 in Tab. IV), so without variance estimates the claimed improvements cannot be distinguished from noise. Please report multiple seeds with standard deviations or a paired significance test on the main comparisons.
  4. [Abstract and Sec. IV-C] The headline numbers are mutually inconsistent. The abstract claims 'at least 2.24%' on ScanRefer, but the Multiple@0.25 gap in Tab. I is 2.04 points; Sec. IV-C says UniSpace-3D surpasses EDA by 3.2% Acc@0.5, whereas the table difference is 2.24 points, and it says Nr3D improves by 5.1% while Tab. II shows a 5.3-point gap. The same section reports '57.8%' on Nr3D, while Tab. II lists 57.2%. These discrepancies must be reconciled before the headline claim is credible.
  5. [Eqs. (1), (2), and (9)] The symbol L_pos is used for two different losses: the positive contrastive loss in Eq. (2) and the position loss in Eq. (9). The final objective in Eq. (9) is therefore ambiguous and cannot be implemented as written. Please rename one of the losses (for example, use L_align and L_bbox) and state the exact total loss with all weight factors.
  6. [Sec. III-B4] The construction of negative grounding sentences with GPT-3 is not sufficiently specified. The paper does not give the prompt used, the number of negative sentences generated per positive, or the criteria for the manual removal step, and it is unclear whether the same negatives are reused across training epochs. Because the MMCL ablation depends on this negative mining, these details are needed to reproduce the method and to rule out accidental dataset leakage.
minor comments (5)
  1. [Tab. IV] The checkmark pattern in Tab. IV is difficult to parse; for example, row (d) is described in the text as the LGQS experiment, but the printed grid appears to show only the URE column selected. Please reformat the table so the selected components match the textual description for every row.
  2. [Sec. IV-B and Tab. II] The relationship between the two-stage and single-stage variants needs clarification; the statement that single-stage methods are excluded from Nr3D/Sr3D because ground truth boxes are provided is unclear, especially since both variants are reported for ScanRefer in Tab. I.
  3. [Eq. (8)] The notation 'Max(−1)' is unusual and is not defined before use; please state explicitly that it takes the maximum along the last dimension of the similarity matrix.
  4. [Throughout] The acronym 'UniSpace-3D' is written inconsistently as 'Unispace-3D' in the Abstract, Sec. IV-E, and the Conclusion; please use a single spelling.
  5. [Sec. I, references [4], [5]] References [4] and [5] concern heterogeneous network crawling and trajectory similarity search and do not appear to relate to point-cloud applications; please re-check the citation context.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical benchmark improvements over external baselines, with no self-citation chain and no definitional equivalence between inputs and outputs.

full rationale

The paper does not contain a derivation that reduces to its own inputs. The central claim is that UniSpace-3D outperforms prior methods on ScanRefer and Nr3D/Sr3D, and this is supported by benchmark evaluations against reproduced baselines such as EDA. The unified representation space is implemented by feeding tokenized point-cloud patches and text through a frozen CLIP encoder; the CLIP pretraining is an external, fixed resource and the grounding accuracy is measured on datasets CLIP was not trained on, so the result is not forced by construction. The losses in Eqs. (1)-(9) are standard training objectives that align embeddings, but the reported Acc@IoU numbers are computed on held-out annotations and are not algebraic consequences of the loss definitions. The closest issue is that the URE ablation (Tab. IV) jointly adds learnable MLPs with the frozen CLIP transformer, so the isolated contribution of CLIP is not separately verified; however, this is an attribution or experimental-design concern, not circularity. There are no self-citations, no imported uniqueness theorems, and no fitted parameter that is later renamed as a prediction. Hyperparameters such as alpha, beta, and gamma are tuned on ScanRefer validation and transferred to other datasets, which is a generalization risk rather than a circular step. Overall, the paper's evidence is empirical and externally grounded, so the appropriate circularity score is 0.

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

The central method rests on the transferability of CLIP's 2D image-text space to 3D point clouds, plus several hand-chosen hyperparameters (loss weights, temperature, query count). No new physical or mathematical entities are introduced; the 'unified representation space' is a learned CLIP embedding space, not a separately postulated entity.

free parameters (3)
  • Loss weights alpha, beta, gamma in Eq. (9) = alpha=0.5, beta=0.3, gamma=0.1
    Selected by grid search on the ScanRefer validation set (Sec. IV-D1, Tab. III) and reused for all reported results.
  • Temperature tau in contrastive losses (Eqs. 3-4) = Not reported
    Scales logits in the MMCL positive contrastive loss; value is chosen by hand but not stated.
  • Number of object candidate queries Nq = 256
    Set in Sec. III-C; affects the decoder's candidate set and is not swept.
assumptions (3)
  • domain assumption Local point cloud patches can be treated as 2D manifold tokens and processed by a frozen CLIP image transformer trained on RGB images.
    Invoked in Sec. III-A3; the entire URE depends on CLIP's image transformer accepting point-cloud-derived tokens.
  • domain assumption CLIP's pre-trained text-image embedding space is a suitable unified space for 3D point-cloud and text alignment after the proposed tokenization.
    Sec. III-A; if CLIP space does not transfer to 3D geometry, the 'unified representation space' would not reduce the modality gap.
  • domain assumption GPT-3-generated negative sentences, after manual filtering, are plausible hard negatives that improve contrastive learning without introducing bias.
    Sec. III-B4; the manual removal step assumes human judgment about scene fit does not distort training.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unified Representation Space for 3D Visual Grounding." pith.science (2026). https://pith.science/paper/YFPNZ7ES

@misc{pith2026250614238,
  author       = {Pith},
  title        = {Pith review of: Unified Representation Space for 3D Visual Grounding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YFPNZ7ES}},
  note         = {Machine review of arXiv:2506.14238}
}
read the original abstract

3D visual grounding (3DVG) is a critical task in scene understanding that aims to identify objects in 3D scenes based on text descriptions. However, existing methods rely on separately pre-trained vision and text encoders, resulting in a significant gap between the two modalities in terms of spatial geometry and semantic categories. This discrepancy often causes errors in object positioning and classification. The paper proposes UniSpace-3D, which innovatively introduces a unified representation space for 3DVG, effectively bridging the gap between visual and textual features. Specifically, UniSpace-3D incorporates three innovative designs: i) a unified representation encoder that leverages the pre-trained CLIP model to map visual and textual features into a unified representation space, effectively bridging the gap between the two modalities; ii) a multi-modal contrastive learning module that further reduces the modality gap; iii) a language-guided query selection module that utilizes the positional and semantic information to identify object candidate points aligned with textual descriptions. Extensive experiments demonstrate that UniSpace-3D outperforms baseline models by at least 2.24% on the ScanRefer and Nr3D/Sr3D datasets. The code will be made available upon acceptance of the paper.

Figures

Figures reproduced from arXiv: 2506.14238 by the authors.

Figure 1
Figure 1. Comparison between the prior works (a) and ours (b). Our method achieves more accurate grounding by bridging the gap between the visual and textual features into the unified representation (UR) space. As illustrated on the right, mapping text and point clouds into the UR space enhances cross￾model correlation, with yellow indicating stronger alignment. localize objects through language-guided keypoint detection, hav… view at source ↗
Figure 2
Figure 2. Overview of UniSpace-3D. The framework comprises three key components. (a) URE, which maps features from point clouds and text descriptions into a unified representation space. (b) MMCL, which refines alignment by reducing the gap between visual and textual embeddings. (c) LGQS, which identifies object candidate points that closely correspond to the text descriptions, enhancing grounding accuracy. A. Unified Represe… view at source ↗
Figure 3
Figure 3. Negative contrastive learning in Multi-Modal Contrastive Learning module. MMCL encourages higher compatibility scores between the true grounding scene and the corresponding sentence while discouraging mismatched pairs. higher cross-correlation, indicating improved alignment within the same scene. B. Multi-Modal Contrastive Learning After mapping the visual and text tokens into the UR space, we aim to minimize the re… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of grounding results from different models on the ScanRefer dataset. Green boxes represent ground-truth references. Red boxes show EDA results containing grounding errors (e.g., objects of the same category as the target). Blue boxes represent proposals g…
Figure 5
Figure 5. Figure 5: Qualitative comparison of the grounding results in the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of some common failure cases. Green boxes represent ground-truth references. Blue boxes represent proposals generated by ours. V. CONCLUSION This paper introduces UniSpace-3D, a unified representation space for 3D visual grounding. UniSpace-3D lever…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages

  1. [1]

    Return to antikythera: Multi-session SLAM based AUV mapping of a first century B.C. wreck site,

    S. B. Williams, O. Pizarro, and B. Foley, “Return to antikythera: Multi-session SLAM based AUV mapping of a first century B.C. wreck site,” inField and Service Robotics - Results of the 10th International Conference, Toronto, Canada, 23-26 June 2015, ser. Springer Tracts in Advanced Robotics, D. Wettergreen and T. D. Barfoot, Eds., vol. 113. Springer, 201...

    doi:10.1007/978-3-319-27702-8 2015

  2. [2]

    Towards subjective quality assessment of point cloud imaging in augmented reality,

    E. Alexiou, E. Upenik, and T. Ebrahimi, “Towards subjective quality assessment of point cloud imaging in augmented reality,” in19th IEEE International Workshop on Multimedia Signal Processing, MMSP 2017, Luton, United Kingdom, October 16-18, 2017. IEEE, 2017, pp. 1–6. [Online]. Available: https://doi.org/10.1109/MMSP.2017.8122237

    work page arXiv 2017

  3. [3]

    Deep hough voting for 3d object detection in point clouds,

    C. R. Qi, O. Litany, K. He, and L. J. Guibas, “Deep hough voting for 3d object detection in point clouds,” in2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019. IEEE, 2019, pp. 9276–9285. [Online]. Available: https://doi.org/10.1109/ICCV .2019.00937

    arXiv 2019

  4. [4]

    Het- erogeneous network crawling: Reaching target nodes by motif-guided navigation,

    C. Wang, K. C.-C. Chang, P. Wang, T. Qin, and X. Guan, “Het- erogeneous network crawling: Reaching target nodes by motif-guided navigation,”IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 9, pp. 4285–4297, 2020

    work page 2020

  5. [5]

    Continuous trajectory similarity search for online outlier detection,

    D. Zhang, Z. Chang, S. Wu, Y . Yuan, K.-L. Tan, and G. Chen, “Continuous trajectory similarity search for online outlier detection,” IEEE Transactions on Knowledge and Data Engineering, vol. 34, no. 10, pp. 4690–4704, 2020

    work page 2020

  6. [6]

    A human-computer interactive method for projected clustering,

    C. C. Aggarwal, “A human-computer interactive method for projected clustering,”IEEE transactions on knowledge and data engineering, vol. 16, no. 4, pp. 448–460, 2004. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 9

    work page 2004

  7. [7]

    Tree-based mining for discovering patterns of human interaction in meetings,

    Z. Yu, Z. Yu, X. Zhou, C. Becker, and Y . Nakamura, “Tree-based mining for discovering patterns of human interaction in meetings,”IEEE Transactions on Knowledge and Data Engineering, vol. 24, no. 4, pp. 759–768, 2010

    work page 2010

  8. [8]

    Scanrefer: 3d object local- ization in rgb-d scans using natural language,

    D. Z. Chen, A. X. Chang, and M. Nießner, “Scanrefer: 3d object local- ization in rgb-d scans using natural language,” inEuropean conference on computer vision, 2020, pp. 202–221

    work page 2020

Show all 39 references
  1. [9]

    3dvg-transformer: Relation modeling for visual grounding on point clouds,

    L. Zhao, D. Cai, L. Sheng, and D. Xu, “3dvg-transformer: Relation modeling for visual grounding on point clouds,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2928–2937

    2021

  2. [10]

    Sat: 2d semantics assisted training for 3d visual grounding,

    Z. Yang, S. Zhang, L. Wang, and J. Luo, “Sat: 2d semantics assisted training for 3d visual grounding,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 1856–1866

    2021

  3. [11]

    3d-sps: Single-stage 3d visual grounding via referred point progressive selection,

    J. Luo, J. Fu, X. Kong, C. Gao, H. Ren, H. Shen, H. Xia, and S. Liu, “3d-sps: Single-stage 3d visual grounding via referred point progressive selection,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 16 454–16 463

    2022

  4. [12]

    Viewpoint-aware visual grounding in 3d scenes,

    X. Shi, Z. Wu, and S. Lee, “Viewpoint-aware visual grounding in 3d scenes,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14 056–14 065

    2024

  5. [13]

    Eda: Explicit text-decoupling and dense alignment for 3d visual grounding,

    Y . Wu, X. Cheng, R. Zhang, Z. Cheng, and J. Zhang, “Eda: Explicit text-decoupling and dense alignment for 3d visual grounding,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19 231–19 242

    2023

  6. [14]

    Scan2cap: Context-aware dense captioning in RGB-D scans,

    D. Z. Chen, A. Gholami, M. Nießner, and A. X. Chang, “Scan2cap: Context-aware dense captioning in RGB-D scans,” inIEEE Conference on Computer Vision and Pattern Recognition, 2021, pp. 3193–3203

    2021

  7. [15]

    X- trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning,

    Z. Yuan, X. Yan, Y . Liao, Y . Guo, G. Li, S. Cui, and Z. Li, “X- trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 8553–8563

    2022

  8. [16]

    End-to-end 3d dense captioning with vote2cap-detr,

    S. Chen, H. Zhu, X. Chen, Y . Lei, G. Yu, and T. Chen, “End-to-end 3d dense captioning with vote2cap-detr,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 11 124–11 133

    2023

  9. [17]

    Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes,

    P. Achlioptas, A. Abdelreheem, F. Xia, M. Elhoseiny, and L. J. Guibas, “Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes,” inComputer Vision - ECCV 2020 - 16th European Conference, ser. Lecture Notes in Computer Science, vol. 12346, 20...

    2020

  10. [18]

    Bottom up top down detection transformers for language grounding in images and point clouds,

    A. Jain, N. Gkanatsios, I. Mediratta, and K. Fragkiadaki, “Bottom up top down detection transformers for language grounding in images and point clouds,” inEuropean Conference on Computer Vision, 2022, pp. 417–433

    2022

  11. [19]

    Scanqa: 3d question answering for spatial scene understanding,

    D. Azuma, T. Miyanishi, S. Kurita, and M. Kawanabe, “Scanqa: 3d question answering for spatial scene understanding,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 19 107–19 117

    2022

  12. [20]

    SQA3D: situated question answering in 3d scenes,

    X. Ma, S. Yong, Z. Zheng, Q. Li, Y . Liang, S. Zhu, and S. Huang, “SQA3D: situated question answering in 3d scenes,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  13. [21]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. A. Funkhouser, and M. Nießner, “Scannet: Richly-annotated 3d reconstructions of indoor scenes,” in2017 IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 2432–2443

    2017

  14. [22]

    Multi-view transformer for 3d visual grounding,

    S. Huang, Y . Chen, J. Jia, and L. Wang, “Multi-view transformer for 3d visual grounding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 15 524–15 533

    2022

  15. [23]

    Secg: Semantic-enhanced 3d visual grounding via cross-modal graph attention,

    F. Xiao, H. Xu, Q. Wu, and W. Kang, “Secg: Semantic-enhanced 3d visual grounding via cross-modal graph attention,”CoRR, vol. abs/2403.08182, 2024. [Online]. Available: https://doi.org/10.48550/ arXiv.2403.08182

    2024 doi arXiv

  16. [24]

    3drp-net: 3d relative position-aware network for 3d visual grounding,

    Z. Wang, H. Huang, Y . Zhao, L. Li, X. Cheng, Y . Zhu, A. Yin, and Z. Zhao, “3drp-net: 3d relative position-aware network for 3d visual grounding,”arXiv preprint arXiv:2307.13363, 2023

    2023 arXiv

  17. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, 2017, pp. 5998–6008

    2017

  18. [26]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” inProceedings of the 38th International Conference on Machine ...

    2021

  19. [27]

    Roberta: A robustly optimized bert pretraining approach,

    Y . Liu, “Roberta: A robustly optimized bert pretraining approach,”arXiv preprint arXiv:1907.11692, vol. 364, 2019

    1907 arXiv

  20. [28]

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

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 652–660

    2017

  21. [29]

    Group-free 3d object detection via transformers,

    Z. Liu, Z. Zhang, Y . Cao, H. Hu, and X. Tong, “Group-free 3d object detection via transformers,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2949–2958

    2021

  22. [30]

    Epcl: Frozen clip transformer is an efficient point cloud encoder,

    X. Huang, Z. Huang, S. Li, W. Qu, T. He, Y . Hou, Y . Zuo, and W. Ouyang, “Epcl: Frozen clip transformer is an efficient point cloud encoder,”arXiv e-prints, pp. arXiv–2212, 2022

    2022

  23. [31]

    P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks,

    X. Liu, K. Ji, Y . Fu, W. Tam, Z. Du, Z. Yang, and J. Tang, “P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks,”Cornell University - arXiv,Cornell University - arXiv, 2021

    2021

  24. [32]

    Contrastive learning for weakly supervised phrase grounding,

    T. Gupta, A. Vahdat, G. Chechik, X. Yang, J. Kautz, and D. Hoiem, “Contrastive learning for weakly supervised phrase grounding,” in European Conference on Computer Vision, 2020, pp. 752–768

    2020

  25. [33]

    Language mod- els are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language mod- els are few-shot learners,”Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  26. [34]

    Grounding dino: Marrying dino with grounded pre- training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, C. Li, J. Yang, H. Su, J. Zhuet al., “Grounding dino: Marrying dino with grounded pre- training for open-set object detection,”arXiv preprint arXiv:2303.05499, 2023

    2023 arXiv

  27. [35]

    Text-guided graph neural networks for referring 3d instance segmentation,

    P. Huang, H. Lee, H. Chen, and T. Liu, “Text-guided graph neural networks for referring 3d instance segmentation,” inThirty-Fifth AAAI Conference on Artificial Intelligence, 2021, pp. 1610–1618

    2021

  28. [36]

    Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring,

    Z. Yuan, X. Yan, Y . Liao, R. Zhang, S. Wang, Z. Li, and S. Cui, “Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring,” in 2021 IEEE/CVF International Conference on Computer Vision, 2021, pp. ...

    2021

  29. [37]

    Free-form description guided 3d visual graph network for object grounding in point cloud,

    M. Feng, Z. Li, Q. Li, L. Zhang, X. Zhang, G. Zhu, H. Zhang, Y . Wang, and A. Mian, “Free-form description guided 3d visual graph network for object grounding in point cloud,” in2021 IEEE/CVF International Conference on Computer Vision, 2021, pp. 3702–3711

    2021

  30. [38]

    Language conditioned spatial relation reasoning for 3d object grounding,

    S. Chen, P. Guhur, M. Tapaswi, C. Schmid, and I. Laptev, “Language conditioned spatial relation reasoning for 3d object grounding,” in Advances in Neural Information Processing Systems, 2022

    2022

  31. [39]

    Viewrefer: Grasp the multi-view knowledge for 3d visual grounding,

    Z. Guo, Y . Tang, R. Zhang, D. Wang, Z. Wang, B. Zhao, and X. Li, “Viewrefer: Grasp the multi-view knowledge for 3d visual grounding,” inIEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, 2023, pp. 15 326–15 337

    2023

Pith tools

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