Pith. sign in

REVIEW 4 major objections 5 minor 61 references

SeqVLM: Proposal-Guided Multi-View Sequences Reasoning via VLM for Zero-Shot 3D Visual Grounding

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SeqVLM claims zero-shot 3D grounding hits 55.6% accuracy via multi-view VLM reasoning, beating prior zero-shot methods by 4–5 points.

desk verdict Sensible pipeline and a solid controlled-VLM comparison, but the SOTA claim rests on an undisclosed 250-sample subset and gains are within sampling error. read the letter →

arxiv 2508.20758 v1 pith:MUIOCX3X submitted 2025-08-28 cs.CV cs.AI

Jiawen Lin , Shiran Bian , Yihang Zhu , Wenbin Tan , Yachao Zhang , Yuan Xie , Yanyun Qu This is my paper
classification cs.CVcs.AI
keywords 3Dvisualgroundingzero-shotlearningvision-languagemodelmulti-viewprojectionsemanticsegmentationproposalsiterativereasoningScanReferNr3D
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

SeqVLM aims to show that a zero-shot 3D visual grounding system can reach near-supervised accuracy by turning a 3D scene into a small set of richly annotated multi-view image sequences and letting a pretrained vision-language model (VLM) reason over them. The pipeline first extracts object proposals with a supervised 3D semantic segmentation network, filters them by category similarity to the query using a CLIP text encoder, then projects each surviving proposal onto five real camera views, stitching the annotated views into one vertical image sequence. An iterative scheduling mechanism feeds these sequences to a VLM in small batches, progressively eliminating candidates until one target remains. On the ScanRefer and Nr3D benchmarks, the paper reports Acc@0.25 of 55.6% and 53.2%, respectively, exceeding prior zero-shot methods by 4.0 and 5.2 points, and matching several fully supervised models on specific splits. The significance is that 3D grounding may not need task-specific training if a good proposal generator and a strong generic VLM are combined through a geometry-aware 2D projection.

What carries the argument

The load-bearing mechanism is the proposal-guided multi-view projection: each 3D instance proposal from Mask3D is projected through the camera intrinsic/extrinsic matrices into each real scene image, validated by a depth-consistency test (Eq. 9), and its bounding rectangle is expanded by a fixed ratio and marked with a red border; the top-five largest-area views are vertically concatenated into one image sequence per proposal. This representation is what lets a generic VLM compare multiple candidates across viewpoints, and the iterative reasoning algorithm (Algorithm 1) is what keeps that comparison within VLM input limits by eliminating candidates batch by batch.

What would settle it

Run SeqVLM and the comparison methods on the full ScanRefer and Nr3D validation sets (or on a much larger random subset with multiple seeds) and check whether the Acc@0.25 margins of 4.0 and 5.2 points persist; also verify that the identical 250 samples are used across all methods. If the gains shrink or reverse on the full set, the central claim about multi-view representation is not established.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the bottleneck in zero-shot 3D visual grounding is not the VLM's language understanding but the form in which the 3D scene is presented to it. Earlier VLM-based methods either render a single synthetic view, which loses occluded detail and real-world texture, or project 2D detections back into 3D without geometric constraints. SeqVLM instead projects 3D instance proposals onto multiple real camera images with a depth-consistency check, annotates each proposal's 2D footprint with a red rectangle, and stitches the top five views into a single vertical image sequence. Fed these sequences plus the textual query, a VLM performs iterative batch elimination unt

Load-bearing premise

The reported 4.0–5.2 point gains rest on a 250-sample validation subset per benchmark whose selection rule is not disclosed; if that subset favors feasible queries or differs from the full benchmark composition, the SOTA claim may not transfer.

Editorial extensions

If this is right

  • If correct, zero-shot 3D grounding can be assembled from off-the-shelf components (a supervised segmentation network, CLIP, an LLM, and a VLM) with no scene-specific training.
  • The 4–5 point gains over prior zero-shot methods imply that multi-view real-image context is more informative for VLM reasoning than single synthetic renders.
  • The framework's accuracy approaches fully supervised methods on several ScanRefer and Nr3D splits, suggesting the supervision bottleneck for 3DVG may shift from grounding models to proposal generators.
  • The iterative batch scheduling makes VLM-based grounding cost-controllable, trading more rounds of inference for the ability to handle many candidates.

Reading between the lines

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

  • The paper's reliance on Mask3D, a supervised segmentation network trained on 3D instance data, means the "zero-shot" claim is really about grounding without paired text-scene supervision; if a fully open-vocabulary 3D proposal generator emerges, the same multi-view projection and iterative reasoning could be retargeted with minimal change.
  • Because the evaluation is on 250 fixed samples per benchmark, the reported SOTA margin may change if the subset is not representative; a natural test is to report accuracy on the full validation set or on multiple random subsets with variance.
  • The red-box annotation and vertical stitching are a simple but strong inductive bias; one could test whether removing the depth-consistency check or varying the expansion ratio alpha changes VLM accuracy, isolating where the geometric projection actually helps.
  • The iterative elimination could be extended to return a ranked list instead of a single winner, which would let downstream systems request the next-best object if the top match is rejected by another modality.
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 / 5 minor

Summary. The paper proposes SeqVLM, a zero-shot 3D visual grounding framework that combines 3D instance proposals from Mask3D, CLIP-based semantic filtering, projection of proposals onto multi-view real images, and iterative VLM selection. The contributions are a proposal-guided multi-view projection module and an iterative reasoning mechanism. Experiments on ScanRefer and Nr3D report Acc@0.25 of 55.6% and 53.2%, claiming state-of-the-art zero-shot performance and competitiveness with fully supervised methods. The evaluation follows a 250-validation-sample protocol per benchmark, as stated in Section 4.1.

Significance. If the reported numbers are reliable, the paper makes a useful architectural contribution: it shows that a pipeline of supervised 3D segmentation, CLIP category filtering, multi-view red-box projection, and iterative VLM selection can perform zero-shot 3D grounding without 3D-grounding-specific training. The controlled cross-method comparison in Table 5, where the VLM backend is held fixed, is a genuine attempt to attribute gains to the framework rather than the VLM. The projection equations in Section 3.4 are standard geometry, and no circular fitting of benchmark numbers to parameters is evident. The code is released, which is a practical strength. However, the central SOTA and 'rivals supervised' claims rest entirely on a 250-sample evaluation whose composition, overlap, and variance are not disclosed; this is the main weakness.

major comments (4)
  1. [Section 4.1, Tables 1-2] The paper states that evaluation uses 250 validation samples per benchmark, following VLM-Grounder, but does not disclose how these 250 samples are selected, whether they are the same 250 samples for every compared method, or any measure of variance. With n=250, the standard error of an accuracy near 50% is about 3.2 percentage points. The headline gains of 4.0% (ScanRefer) and 5.2% (Nr3D) are only about 1.2-1.6 standard errors, and per-cell margins in Table 1 (e.g., Unique Acc@0.25 +1.6, Multiple Acc@0.25 -0.5) are much smaller. The claim that SeqVLM establishes a new state of the art is therefore not statistically supported as reported. The authors should either evaluate on the full validation sets or provide a detailed sampling protocol, identical sample indices across methods, and bootstrap confidence intervals.
  2. [Section 4.4, Figures 5-6, Table 4] The hyperparameters L (batch size threshold) and n_frame (multi-view frame number), as well as the VLM choice (Doubao-1.5-vision-pro), are selected using the same 250-sample evaluation set on which the final numbers are reported. This introduces selection bias: the reported values may be optimized for this particular subsample rather than reflecting the general behavior of the method. A separate validation split or a sensitivity analysis over multiple subsamples is needed before the reported Acc values can be treated as unbiased estimates.
  3. [Section 4.4, Table 5] The controlled-VLM comparison is the strongest evidence for the architectural contribution, but it only reports Acc@0.5, not the headline Acc@0.25 used in the SOTA claim. Moreover, the table does not explicitly state that all four methods are evaluated on exactly the same 250 samples. Since the proposal and projection pipelines differ across methods, controlling only the VLM backend does not fully control for sample composition or for differences in the number/quality of candidate proposals. The paper should state the exact sample indices and, ideally, report both Acc@0.25 and Acc@0.5 for all methods under identical subsampling.
  4. [Tables 1-2 and Section 4.2] The comparison with fully supervised methods is apples-to-oranges. The fully supervised numbers in Tables 1 and 2 are taken from the original papers and are typically computed on the full validation set, whereas SeqVLM is evaluated on a 250-sample subset. If the subset is not representative, the claim that SeqVLM 'rivals fully supervised approaches' (Introduction and Section 5) is not established. This comparison should either use the full validation set or be explicitly framed as a preliminary subset-based comparison.
minor comments (5)
  1. [Algorithm 1] The title has a typo: 'Interative' should be 'Iterative'. Also, the final return when Q becomes empty is None, but the main text does not discuss how this case is handled in the overall pipeline; please clarify the behavior when no candidate survives an iteration.
  2. [Section 3.4, Eq. (9)] The depth consistency check divides by D(u,v), which can be zero or missing in real depth images. Please specify the handling of invalid depth values before applying the threshold tau.
  3. [Throughout] The naming of the baseline is inconsistent: 'ZS3DVG' in Tables 1 and 2 versus 'ZSVG3D' in the text and Table 5. Please unify.
  4. [References] Reference [34] (MCLN) is missing venue and year details; Reference [3] points to a general VolcEngine console URL, which makes the exact model version and API settings hard to reproduce. Please include versioned identifiers or commit hashes for closed APIs.
  5. [Figure 5] The axis label 'VLM Capacity Threshold L' is confusing because the x-axis shows L values while the caption says threshold L. Please relabel to 'Batch size threshold L' for consistency with Section 3.5.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SeqVLM's pipeline is a composition of external pretrained components and standard geometric projections; the reported Acc numbers are empirical evaluations, not quantities forced by the method's definitions.

full rationale

The derivation chain is self-contained as an engineering system: Mask3D proposals (external, trained on ScanNet), CLIP text-category filtering (external), pinhole projection with depth verification (standard geometry, Eqs. 7-9), multi-view stitching, and iterative VLM selection via the Doubao API. None of these steps defines the target answer in terms of the final benchmark score. The final accuracy can fail in multiple non-degenerate ways (proposal misses the target, category filter drops it, projection occludes it, or VLM chooses the wrong index), so the result is not true by construction. No load-bearing self-citation appears: the paper's cited prior works are by other groups, and the 'standardized protocol' of VLM-Grounder cited in Sec. 4.1 is an external evaluation convention, not a self-citation chain. The 250-sample evaluation protocol is a genuine external-validity limitation (small n, undisclosed sample selection, shared-sample ambiguity), and the hyperparameter ablations in Figs. 5-6 are tuned on that same subset; these are statistical/selection concerns about whether the SOTA margin generalizes, not circularity in the sense of an equation reducing to its inputs or a fitted parameter renamed as a prediction. The paper's open admission of the 250-sample protocol is weighed here as a non-circular limitation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper contributes a pipeline; its parameters are hand-set or tuned on the evaluation subsets. The main 'input' from prior work is a supervised 3D segmenter trained on the same dataset family, plus a closed-source VLM; no new physical or conceptual entities are introduced.

free parameters (6)
  • segmentation confidence threshold theta = 0.2
    Hand-set threshold to filter Mask3D masks (Eq. 1, Implementation Details).
  • visibility threshold tau = 0.25
    Depth consistency tolerance in Eq. 9; hand-set.
  • bounding box expansion alpha = 0.25
    Expansion ratio for the red annotation box around each projected proposal (Eq. 10 context); hand-set.
  • batch size threshold L = 4
    VLM batch capacity in Algorithm 1; tuned on ScanRefer via ablation (Figure 5).
  • multi-view frame number n_frame = 5
    Number of top-projected views per proposal; tuned on ScanRefer via ablation (Figure 6).
  • viewpoint sampling interval = 20 frames
    Scene images segmented into 20-frame intervals for viewpoint sampling; hand-set (Implementation Details).
assumptions (4)
  • domain assumption Mask3D predicts reliable instance masks and categories on ScanNet scenes.
    The entire proposal set comes from this supervised 3D segmenter; its failures (missed objects, bad masks) are not evaluated separately. Mask3D is trained on ScanNet data, so this component is not zero-shot (Section 3.3).
  • domain assumption ScanNet camera poses and intrinsics used in Eqs. 7-8 are accurate.
    Projection validity and depth consistency (Eq. 9) assume registered RGB-D frames; no error analysis on pose calibration or frame-to-mesh alignment is given (Section 3.4).
  • domain assumption The closed Doubao-1.5-vision-pro VLM returns reliable selection indices for the constructed prompts.
    VLM output is the final decision; prompt sensitivity, API version drift, and non-determinism are not controlled (Section 3.5, Tables 4-5).
  • domain assumption The 250-sample validation subsets are representative of the full ScanRefer and Nr3D benchmarks.
    Headline results are computed on these subsets only (Implementation Details); no subset description, selection rule, or confidence intervals are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SeqVLM: Proposal-Guided Multi-View Sequences Reasoning via VLM for Zero-Shot 3D Visual Grounding." pith.science (2026). https://pith.science/paper/MUIOCX3X

@misc{pith2026250820758,
  author       = {Pith},
  title        = {Pith review of: SeqVLM: Proposal-Guided Multi-View Sequences Reasoning via VLM for Zero-Shot 3D Visual Grounding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MUIOCX3X}},
  note         = {Machine review of arXiv:2508.20758}
}
read the original abstract

3D Visual Grounding (3DVG) aims to localize objects in 3D scenes using natural language descriptions. Although supervised methods achieve higher accuracy in constrained settings, zero-shot 3DVG holds greater promise for real-world applications since eliminating scene-specific training requirements. However, existing zero-shot methods face challenges of spatial-limited reasoning due to reliance on single-view localization, and contextual omissions or detail degradation. To address these issues, we propose SeqVLM, a novel zero-shot 3DVG framework that leverages multi-view real-world scene images with spatial information for target object reasoning. Specifically, SeqVLM first generates 3D instance proposals via a 3D semantic segmentation network and refines them through semantic filtering, retaining only semantic-relevant candidates. A proposal-guided multi-view projection strategy then projects these candidate proposals onto real scene image sequences, preserving spatial relationships and contextual details in the conversion process of 3D point cloud to images. Furthermore, to mitigate VLM computational overload, we implement a dynamic scheduling mechanism that iteratively processes sequances-query prompts, leveraging VLM's cross-modal reasoning capabilities to identify textually specified objects. Experiments on the ScanRefer and Nr3D benchmarks demonstrate state-of-the-art performance, achieving Acc@0.25 scores of 55.6% and 53.2%, surpassing previous zero-shot methods by 4.0% and 5.2%, respectively, which advance 3DVG toward greater generalization and real-world applicability. The code is available at https://github.com/JiawLin/SeqVLM.

Figures

Figures reproduced from arXiv: 2508.20758 by the authors.

Figure 1
Figure 1. Comparative overview of previous VLM-based meth [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of SeqVLM. Given a 3D scene and a textual query, SeqVLM localizes the target object through a structured [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results of 3D visual grounding on ScanRefer[ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparative visualization of single-view rendering [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Ablation study on different values of batch size [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Ablation study on different values of multi-view [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 51 canonical work pages

  1. [1]

    Referit3d: Neural listeners for fine-grained 3d object iden- tification in real-world scenes

    Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object iden- tification in real-world scenes. In European Conference on Computer Vision , pages 422–440. Springer, 2020

    work page 2020

  2. [2]

    Tongyi Qianwen

    Alibaba Cloud. Tongyi Qianwen. https://bailian.console.aliyun.com/, 2025. Ac- cessed: 2025-03-03

    work page 2025

  3. [3]

    VolcEngine

    ByteDance. VolcEngine. https://console.volcengine.com/, 2025. Accessed: 2025- 02-18

    work page 2025

  4. [4]

    3djcg: A uni- fied framework for joint dense captioning and visual grounding on 3d point clouds

    Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A uni- fied 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, pages 16464–16473, 2022

    work page 2022

  5. [5]

    Tparn: A network for enhancing synthetic video quality after 3d-hevc encoding

    Ziyi Cao, Tiansong Li, Shaoguo Cui, Kejun Wu, Yan Chen, Longwei Zhong, Hongkui Wang, and Li Yu. Tparn: A network for enhancing synthetic video quality after 3d-hevc encoding. In 2024 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5. IEEE, 2024

    work page 2024

  6. [6]

    Scanrefer: 3d object localization in rgb-d scans using natural language

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

    work page 2020

  7. [7]

    D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding

    Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D 3 net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In European Conference on Computer Vision , pages 487–505. Springer, 2022

    work page 2022

  8. [8]

    Minigpt-v2: large language model as a unified interface for vision- language multi-task learning

    Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision- language multi-task learning. arXiv preprint arXiv:2310.09478, 2023

    arXiv 2023

Show all 61 references
  1. [9]

    Reasoning3d–grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models

    Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d–grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv prep...

    2024 arXiv

  2. [10]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2023

  3. [11]

    Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field

    Jiarui Hu, Xianhao Chen, Boyin Feng, Guanglin Li, Liangjing Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Cg-slam: Efficient dense rgb-d slam in a consistent uncertainty-aware 3d gaussian field. In European Conference on Computer Vision, pages 93–112. Springer, 2024

    2024

  4. [12]

    Photo-slam: Real- time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras

    Huajian Huang, Longwei Li, Hui Cheng, and Sai-Kit Yeung. Photo-slam: Real- time simultaneous localization and photorealistic mapping for monocular stereo and rgb-d cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21584–21593, 2024

    2024

  5. [13]

    Reason3d: Searching and reasoning 3d segmentation via large language model

    Kuan-Chih Huang, Xiangtai Li, Lu Qi, Shuicheng Yan, and Ming-Hsuan Yang. Reason3d: Searching and reasoning 3d segmentation via large language model. In International Conference on 3D Vision 2025 , 2025

    2025

  6. [14]

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

    Pin-Hao Huang, Han-Hung Lee, Hwann-Tzong Chen, and Tyng-Luh Liu. Text- guided graph neural networks for referring 3d instance segmentation. In AAAI, volume 35, pages 1610–1618, 2021

    2021

  7. [15]

    Training an open-vocabulary monocular 3d detection model without 3d data

    Rui Huang, Henry Zheng, Yan Wang, Zhuofan Xia, Marco Pavone, and Gao Huang. Training an open-vocabulary monocular 3d detection model without 3d data. Advances in Neural Information Processing Systems , 37:72145–72169, 2024

    2024

  8. [16]

    Multi-view transformer for 3d visual grounding

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

    2022

  9. [17]

    Joint semi- supervised and active learning via 3d consistency for 3d object detection

    Sihwan Hwang, Sanmin Kim, Youngseok Kim, and Dongsuk Kum. Joint semi- supervised and active learning via 3d consistency for 3d object detection. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 4819–4825. IEEE, 2023

    2023

  10. [18]

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

    Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision , pages 417–433. Springer, 2022

    2022

  11. [19]

    Cplip: zero-shot learning for histopathology with comprehensive vision-language alignment

    Sajid Javed, Arif Mahmood, Iyyakutti Iyappan Ganapathi, Fayaz Ali Dharejo, Naoufel Werghi, and Mohammed Bennamoun. Cplip: zero-shot learning for histopathology with comprehensive vision-language alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patter...

    2024

  12. [20]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 4867–4876, 2020

    2020

  13. [21]

    Lerf: Language embedded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 19729–19739, 2023

    2023

  14. [22]

    Rgb-d based visual slam algorithm for indoor crowd environment

    Jianfeng Li, Juan Dai, Zhong Su, and Cui Zhu. Rgb-d based visual slam algorithm for indoor crowd environment. Journal of Intelligent & Robotic Systems , 110(1):27, 2024

    2024

  15. [23]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning , pages 19730–19742. PMLR, 2023

    2023

  16. [24]

    Seeground: See and ground for zero-shot open-vocabulary 3d visual grounding

    Rong Li, Shijie Li, Lingdong Kong, Xulei Yang, and Junwei Liang. Seeground: See and ground for zero-shot open-vocabulary 3d visual grounding. arXiv preprint MM ’25, October 27–31, 2025, Dublin, Ireland Jiawen Lin et al. arXiv:2412.04383, 2024

    2025 arXiv

  17. [25]

    Exploring diversity-based active learning for 3d object detection in autonomous driving

    Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems , 2024

    2024

  18. [26]

    Unimel: A unified framework for multimodal entity linking with large language models

    Qi Liu, Yongyi He, Tong Xu, Defu Lian, Che Liu, Zhi Zheng, and Enhong Chen. Unimel: A unified framework for multimodal entity linking with large language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 1909–1919, 2024

    1909

  19. [27]

    Aligning cyber space with physical world: A comprehensive survey on embodied ai

    Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Aligning cyber space with physical world: A comprehensive survey on embodied ai. arXiv preprint arXiv:2407.06886, 2024

    2024 arXiv

  20. [28]

    Group-free 3d object detection via transformers

    Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2949–2958, 2021

    2021

  21. [29]

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

    Junyu Luo, Jiahui Fu, Xianghao Kong, Chen Gao, Haibing Ren, Hao Shen, Huaxia Xia, and Si Liu. 3d-sps: Single-stage 3d visual grounding via referred point progressive selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 16454–1...

    2022

  22. [30]

    Motion detection methods applied on rgb-d images for vehicle classification on the edge computing

    Kristian Micko and Peter Papcun. Motion detection methods applied on rgb-d images for vehicle classification on the edge computing. IEEE Internet of Things Journal, 2025

    2025

  23. [31]

    OpenAI. ChatGPT. https://chatgpt.com/, 2024. Version: 2024-05-13

    2024

  24. [32]

    Openscene: 3d scene understanding with open vo- cabularies

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

    2023

  25. [33]

    Zeetad: Adapting pretrained vision-language model for zero-shot end-to-end temporal action detection

    Thinh Phan, Khoa Vo, Duy Le, Gianfranco Doretto, Donald Adjeroh, and Ngan Le. Zeetad: Adapting pretrained vision-language model for zero-shot end-to-end temporal action detection. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 7046–7...

    2024

  26. [34]

    Multi-branch collaborative learning network for 3d visual grounding

    Zhipeng Qian, Yiwei Ma, Zhekai Lin, Jiayi Ji, and Xiawu Zheng. Multi-branch collaborative learning network for 3d visual grounding

  27. [35]

    Rgb guided tof imaging system: a survey of deep learning-based methods

    Xin Qiao, Matteo Poggi, Pengchao Deng, Hao Wei, Chenyang Ge, and Stefano Mattoccia. Rgb guided tof imaging system: a survey of deep learning-based methods. International Journal of Computer Vision , 132(11):4954–4991, 2024

    2024

  28. [36]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning , p...

    2021

  29. [37]

    Mask3d: Mask transformer for 3d semantic instance segmentation

    Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA) , pages 8216–8223. IEEE, 2023

    2023

  30. [38]

    Layer depth denoising and completion for structured-light rgb-d cameras

    Ju Shen and Sen-Ching S Cheung. Layer depth denoising and completion for structured-light rgb-d cameras. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages 1187–1194, 2013

    2013

  31. [39]

    Flowcam: Training generalizable 3d radiance fields without camera poses via pixel-aligned scene flow

    Cameron Smith, Yilun Du, Ayush Tewari, and Vincent Sitzmann. Flowcam: Training generalizable 3d radiance fields without camera poses via pixel-aligned scene flow. arXiv preprint arXiv:2306.00180, 2023

    2023 arXiv

  32. [40]

    Four ways to improve verbo-visual fusion for dense 3d visual grounding

    Ozan Unal, Christos Sakaridis, Suman Saha, and Luc Van Gool. Four ways to improve verbo-visual fusion for dense 3d visual grounding. In European Conference on Computer Vision , pages 196–213. Springer, 2024

    2024

  33. [41]

    Soft- group for 3d instance segmentation on point clouds

    Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Soft- group for 3d instance segmentation on point clouds. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2708–2717, 2022

    2022

  34. [42]

    A survey on large language model based autonomous agents

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, et al. A survey on large language model based autonomous agents. Frontiers of Computer Science , 18(6):186345, 2024

    2024

  35. [43]

    Gˆ 3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding

    Yuan Wang, Yali Li, and Shengjin Wang. Gˆ 3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 13917–13926, 2024

    2024

  36. [44]

    Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes

    Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023

    2023 arXiv

  37. [45]

    Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks

    Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Zhe Chen, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. Advances in Neural Information Processing Sy...

    2024

  38. [46]

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

    Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 19231–19242, 2023

    2023

  39. [47]

    M-divo: Multiple tof rgb-d cameras enhanced depth-inertial-visual odometry

    Jie Xu, Wenlu Yu, Song Huang, Shenghai Yuan, Lijun Zhao, Ruifeng Li, and Lihua Xie. M-divo: Multiple tof rgb-d cameras enhanced depth-inertial-visual odometry. IEEE Internet of Things Journal , 2024

    2024

  40. [48]

    Vlm-grounder: A vlm agent for zero-shot 3d visual grounding.arXiv preprint arXiv:2410.13860, 2024

    Runsen Xu, Zhiwei Huang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Vlm-grounder: A vlm agent for zero-shot 3d visual grounding.arXiv preprint arXiv:2410.13860, 2024

    2024 arXiv

  41. [49]

    Multi-scale 3d gaussian splatting for anti-aliased rendering

    Zhiwen Yan, Weng Fei Low, Yu Chen, and Gim Hee Lee. Multi-scale 3d gaussian splatting for anti-aliased rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 20923–20931, 2024

    2024

  42. [50]

    Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent

    Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages 7...

    2024

  43. [51]

    Sat: 2d seman- tics assisted training for 3d visual grounding

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

    2021

  44. [52]

    Visual programming for zero-shot open-vocabulary 3d visual grounding

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

    2024

  45. [53]

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

    Zhihao Yuan, Xu Yan, Yinghong Liao, Ruimao Zhang, Sheng Wang, Zhen Li, and Shuguang Cui. Instancerefer: Cooperative holistic understanding for visual grounding on point clouds through instance multi-level contextual referring. In Proceedings of the IEEE/CVF International Confe...

    2021

  46. [54]

    Vision- language pre-training with object contrastive learning for 3d scene understanding

    Taolin Zhang, Sunan He, Tao Dai, Zhi Wang, Bin Chen, and Shu-Tao Xia. Vision- language pre-training with object contrastive learning for 3d scene understanding. In AAAI, volume 38, pages 7296–7304, 2024

    2024

  47. [55]

    Earthgpt: A universal multi-modal large language model for multi-sensor image compre- hension in remote sensing domain

    Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multi-modal large language model for multi-sensor image compre- hension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 2024

    2024

  48. [56]

    Prototype correlation matching and class-relation reasoning for few-shot medical image segmentation

    Yumin Zhang, Hongliu Li, Yajun Gao, Haoran Duan, Yawen Huang, and Yefeng Zheng. Prototype correlation matching and class-relation reasoning for few-shot medical image segmentation. IEEE Transactions on Medical Imaging , 2024

    2024

  49. [57]

    Towards clip-driven language-free 3d visual grounding via 2d-3d relational enhancement and consistency

    Yuqi Zhang, Han Luo, and Yinjie Lei. Towards clip-driven language-free 3d visual grounding via 2d-3d relational enhancement and consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 13063–13072, 2024

    2024

  50. [58]

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

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

    2021

  51. [59]

    A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms

    Wangbo Zhao, Yizeng Han, Jiasheng Tang, Zhikai Li, Yibing Song, Kai Wang, Zhangyang Wang, and Yang You. A stitch in time saves nine: Small vlm is a precise guidance for accelerating large vlms. arXiv preprint arXiv:2412.03324 , 2024

    2024 arXiv

  52. [60]

    Preventing zero-shot transfer degradation in continual learning of vision- language models

    Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xiangyu Yue, and Yang You. Preventing zero-shot transfer degradation in continual learning of vision- language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19125–19136, 2023

    2023

  53. [61]

    Context-aware 3d object detection from a single image in autonomous driving

    Dingfu Zhou, Xibin Song, Jin Fang, Yuchao Dai, Hongdong Li, and Liangjun Zhang. Context-aware 3d object detection from a single image in autonomous driving. IEEE Transactions on Intelligent Transportation Systems , 23(10):18568– 18580, 2022

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.