Pith. sign in

REVIEW 4 major objections 6 minor 47 references

HQ-OV3D: A High Box Quality Open-World 3D Detection Framework based on Diffision Model

T0 review · 4 major / 6 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 HQ-OV3D introduces a two-stage pipeline—IMCV proposal generation plus a DCIM-based ACA denoiser—that produces higher-quality 3D pseudo-labels for unseen classes, yielding a 7.37% mAP gain over prior state-of-the-art on nuScenes.

desk verdict A serious OV-3D pseudo-label refinement framework whose headline 7.37% gain is not yet supported because Table 1 compares against baselines from different protocols; the controlled numbers in Table 3 show a smaller but plausible gain. read the letter →

arxiv 2508.10935 v2 pith:XM2BH53I submitted 2025-08-12 cs.CV cs.LGcs.RO

Qi Liu , Yabei Li , Hongsong Wang , Lei He This is my paper · ORCID
classification cs.CVcs.LGcs.RO
keywords open-vocabulary3Dobjectdetectionpseudo-labelgenerationdiffusionmodelDDIMLiDAR-camerafusionboundingboxrefinementcross-modalityconsistencynuScenes
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

HQ-OV3D is a framework for open-vocabulary 3D object detection that focuses on the geometric quality of pseudo-labels for novel classes. It claims that existing two-stage pipelines generate noisy 3D boxes because they project 2D VLM detections into point clouds heuristically, and it replaces that with (1) an intra-modality cross-validated proposal generator that uses image-LiDAR consistency to build cleaner initial boxes, and (2) an annotated-class-assisted denoiser that transfers geometric priors from known classes to refine those boxes via DDIM diffusion. The paper reports that training a Transfusion detector on these pseudo-labels improves novel-class mAP by 7.37% over the current state of the art, and that the method works both as a standalone detector and as a plug-in pseudo-label generator.

What carries the argument

The load-bearing machinery is the pairing of the Intra-Modality Cross-Validated (IMCV) Proposal Generator and the Annotated-Class Assisted (ACA) Denoiser. IMCV uses 2D VLM boxes plus SAM masks to extract candidate LiDAR points, computes per-point geometry scores and per-cluster geometry confidence, and merges clusters only when the refitted 3D box's projection IoU improves within a dimension threshold; a Proposal Selector then picks the best box from size/yaw candidates scored by point cloud coverage and projection IoU. ACA treats a proposal as a noisy observation $\hat{b}_t = b_0 + \varepsilon_{\text{sys}}$, learns to predict residual $\Delta b_t = (\Delta c_t, \Delta s_t, \Delta \theta_t)$

What would settle it

Measure the empirical distribution of IMCV proposal errors (center offset, size, yaw) on base classes and compare to the Gaussian noise used in training; if the real error distribution is rejected by a normality test (e.g., heavy tails or range-dependent variance), or if a denoiser trained on the actual error distribution beats the Gaussian-trained denoiser on novel-class refinements by more than noise, the central assumption is falsified. A simpler test: apply the ACA denoiser to proposals whose error statistics are deliberately shifted (e.g., all centers offset by +2m) and check whether refi

Watch

Extended reading notes

Core claim

The central claim is that box quality, not semantic alignment, is the limiting factor in open-vocabulary 3D detection, and that a diffusion-based denoiser conditioned on super-categories can transfer geometric knowledge from annotated base classes to unseen novel classes without any novel-class labels. The IMCV generator addresses depth ambiguity, occlusion, and LiDAR sparsity by scoring point clusters with a geometry confidence, merging clusters only when projected IoU improves, and selecting boxes from candidates guided by category size priors. The ACA denoiser then treats each proposal as a noisy observation of a true box and iteratively removes the residual between proposal and ground tr

Load-bearing premise

The ACA denoiser is trained by adding Gaussian noise to ground-truth boxes, but must correct the systematic, potentially non-Gaussian errors of IMCV proposals; if those errors are strongly anisotropic, range-dependent, or occlusion-correlated, the learned correction may be wrong for novel classes.

Editorial extensions

If this is right

  • If valid, pseudo-label quality is a controllable bottleneck: improving box geometry yields direct mAP gains for unseen classes even without changing the semantic alignment stage.
  • The ACA denoiser demonstrates that base-class bounding-box annotations can be recycled as geometric priors for novel classes via super-category-conditioned diffusion, so novel classes do not need annotations or paired point-cloud-text data.
  • HQ-OV3D works as a drop-in pseudo-label generator: any existing OV-3D detector can be trained on these labels, and the 7.37% improvement was measured with a standard Transfusion detector.
  • The GT2D-ablation result implies headroom: if 2D VLMs improve their box accuracy, the full HQ-OV3D pipeline's pseudo-labels and final detection mAP stand to improve substantially.

Reading between the lines

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

  • The Gaussian-noise surrogate for proposal error is the most fragile assumption; a denoiser trained to invert the actual error distribution of IMCV proposals (e.g., estimated from base-class predictions) might generalize better than one trained on synthetic Gaussian perturbations.
  • Super-category conditioning could be made fully automatic by clustering base-class box statistics instead of relying on GPT-4 size priors; this might improve transfer when novel classes have less typical geometry.
  • The same two-stage recipe—cross-modal proposal generation plus class-prior-conditioned denoising—could extend to indoor scenes or radar-lidar fusion, where the error statistics and occlusion patterns differ.
  • A testable extension is to feed HQ-OV3D pseudo-labels into an alignment-based OV-3D method (e.g., OV-SCAN) and see whether the geometric gains compound with semantic alignment gains.
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 HQ-OV3D, a two-stage open-vocabulary 3D detection pipeline for nuScenes. Stage 1 (IMCV) generates 3D proposals from VLM 2D detections through point-in-mask projection, geometric scoring, greedy cluster merging, and box-candidate selection with GPT-4 dimension priors. Stage 2 (ACA) trains a DDIM-based denoiser on base-class ground-truth boxes with Gaussian noise and super-category conditioning, then refines novel-class proposals; an IoU-aware confidence score is fused with the VLM score. The authors claim that training a Transfusion detector on pseudo-labels generated by HQ-OV3D improves novel-class mAP by 7.37% over state-of-the-art (Table 1).

Significance. If the headline result were reproducible, HQ-OV3D would be a useful contribution: it addresses a real bottleneck (pseudo-label box geometry) in open-vocabulary 3D detection, and it includes a clearly specified proposal generator, a denoising objective, and ablations across three VLM backbones (GLIP/GDINO/GT2D). However, the central empirical claim is internally inconsistent: Table 1 and Table 3 report the same method/VLM configuration with very different per-class APs, and the Table 1 baselines are not shown to be controlled. These issues bear directly on the claimed 7.37% improvement, so the manuscript's central claim is not currently supported. The paper also provides no code and no error bars, which limits reproducibility.

major comments (4)
  1. [Table 1 vs. Table 3] Table 1 reports HQ-OV3D with GLIP as 25.30/29.80/51.20/57.80 (mAP 41.02), while Table 3 reports the same configuration as 14.30/18.30/37.30/22.10 (mAP 23.00) on the same four novel classes. In addition, Table 2's 'HQ-OV3D GLIP' row (10.80/12.10/31.30/21.70) is numerically identical to Table 3's 'HQ-OV3D* GLIP' (without ACA) row, suggesting that Table 2 may omit the ACA denoiser despite the naming. Because the 7.37% claim in the abstract and Main Results is computed solely from Table 1, this inconsistency invalidates the headline number as presented.
  2. [Main Results, Table 1] The Table 1 comparison is not controlled. The same Find n' GLIP baseline appears as 26.17/28.43/34.18/45.83 (mAP 33.65) in Table 1 but as 7.90/9.20/32.10/20.50 (mAP 17.43) in Tables 2 and 3 on the same four novel classes. The text states that the authors' method is evaluated by training a Transfusion detector, but it does not state that Find n', OpenSight, or OV-SCAN were retrained under the same protocol. Without this control, the 7.37% gain may reflect a different detector/training schedule rather than pseudo-label quality.
  3. [Appendix, Table 5] The IoU confidence weight of 0.6 is selected as the value that maximizes the same novel-class mAP used for the paper's headline result. No held-out validation split or selection-corrected estimate is reported. Since hyperparameters such as thresh_dim, DBSCAN eps/min_samples, alpha_pts/alpha_IoU, Nb, and Ntheta are also fixed without sensitivity analysis, the reported mAP is a post-selection optimum and cannot be treated as an unbiased estimate of the method's performance.
  4. [Diffusion-Based Refinement, Eq. (8)/(15)] The ACA denoiser is trained with Gaussian perturbations of ground-truth boxes but is deployed to correct the systematic errors of IMCV proposals. The manuscript acknowledges this mismatch in the text ('the actual error distribution may not strictly follow a gaussian distribution...'), yet supplies no empirical comparison between the training noise and the real proposal-error distribution. Since the denoiser is credited with the 4.02 mAP gain in Table 3, this is a load-bearing assumption. Reporting the empirical error statistics of IMCV proposals and/or training with an empirically matched noise model would be necessary to support the claimed refinement.
minor comments (6)
  1. [Title] 'Diffision' should be 'Diffusion'.
  2. [Main Results] The protocol for Table 1 baselines should be stated explicitly: were OpenSight, OV-SCAN, and Find n' all retrained with the same Transfusion detector, or are the numbers imported from the original papers?
  3. [Table 3 caption] The caption should specify the exact evaluation protocol (detector, training split, and number of classes) so the reader can reconcile Table 3 with Table 1.
  4. [Eq. (10)] Clarify whether z is sampled at inference. If beta_t > 0, the 'deterministic sampling path' claim for DDIM needs qualification.
  5. [Throughout] Typos and wording: 'comprises of', 'V ocabulary', 'pesudo', 'exiting', 'a Annotated', and inconsistent 'Find n'' vs 'Find n' Propagate'.
  6. [Experimental Setup] No code, random seeds, or error bars are provided; given the table inconsistencies, this severely limits reproducibility.

Circularity Check

1 steps flagged · score 4.0 of 10

One scalar confidence weight is tuned on the reported novel-class mAP; the headline baseline comparison is also internally inconsistent, but the core IMCV+ACA pipeline is not circular.

  1. fitted input called prediction [Appendix, Ablation Study of IoU-Aware Confidence Score, Table 5]
    "From Table 5, we observe that incorporating the IoU-Aware Confidence Score leads to a moderate improvement in mAP, with the best performance achieved at a weight of 0.6."

    The IoU/VLM confidence fusion weight is a hyperparameter of the final pseudo-label-ranking step. The paper selects it by sweeping the weight and taking the value that maximizes the exact novel-class mAP reported as evidence (Table 5: mAP 27.78 at weight 0.6 vs 26.88 at 0.0). The reported mAP is therefore the maximum of the selection objective, not an independent prediction of pseudo-label quality. Since the main results use this best-weight configuration, part of the margin over untuned comparisons is a selection artifact. This is a limited, partial circularity; the core IMCV+ACA pipeline itself is evaluated by held-out boxes and is not defined in terms of the metric.

full rationale

The paper's derivation chain is largely self-contained: the IMCV generator uses VLM+SAM proposals and geometric consistency, and the ACA denoiser is trained on base-class boxes with Gaussian perturbations; neither module is defined in terms of the novel-class mAP used for evaluation, and there are no load-bearing self-citations. The Gaussian-noise surrogate for the real proposal error is an empirical assumption (quoted in the Diffusion-Based Refinement subsection, Eq. 8/15 context), not a circularity, because the training labels are base-class GT boxes, not the IMCV proposals being refined at inference. The GPT-4 dimension priors (Table 6) are external world-knowledge inputs, not derived from the target outputs, so their use is not circular. The clearest circularity-adjacent step is the selection of the IoU/VLM confidence weight in Table 5: the weight is chosen by maximizing the same novel-class mAP later reported as evidence of pseudo-label quality, making the reported mAP a fitted statistic for that scalar. Separately, there is a load-bearing correctness concern: Table 1's Find n' baseline (33.65 mAP) is not reproduced under the paper's own controlled protocol in Table 3 (17.43 mAP for the same method, four novel classes, nuScenes); this suggests the headline 7.37% improvement mixes protocols and should be verified against retrained baselines. This issue does not make the core IMCV+ACA derivation circular, but it does weaken the headline empirical claim. Overall circularity score 4: one partial fit-to-evaluation step, with the central pipeline retaining independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central improvement rests on several untested or hand-chosen inputs: Gaussian surrogate noise, GPT-4 priors, hand super categories, and tuning of confidence weights on the evaluation metric. The counts above show that the empirical margin is not a parameter-free result; however, none of these are circular in the strongest sense because the final mAP is still measured against held-out novel class boxes.

free parameters (6)
  • IoU confidence weight = 0.6 (with VLM weight 0.4)
    Selected in Table 5 by sweeping the weight to maximize mAP on the evaluation split; the reported result uses the best-found value.
  • thresh_dim factor = 1.2 x dimension prior
    Hand-set threshold in Algorithm 1 for deciding whether cluster merging is retained; appendix says thresh_dim = 1.2 x dim_prior.
  • DBSCAN eps and min_samples = eps = 0.50, min_samples = 1
    Chosen without sensitivity analysis in Implementation Details; affects cluster formation in Object Localizer.
  • GPT-4 category dimension priors = Table 6 values per category
    Queried from GPT-4 and used to build candidate boxes and thresholds for novel classes; no independent calibration or uncertainty reported.
  • alpha_pts and alpha_IoU = not stated (alpha_pts + alpha_IoU = 1)
    Weights in Eq. 7 for candidate box selection; values not reported, so reproducibility requires reverse engineering.
  • Number of candidate boxes Nb and yaw angles Ntheta = not stated
    Candidate set size in Eq. 5 is unspecified, directly affecting search space and results.
assumptions (5)
  • ad hoc to paper Gaussian noise is a general-purpose surrogate for the true systematic error of proposals.
    Admitted in the ACA Denoiser training paragraph; the denoiser is trained on Gaussian-perturbed GT boxes but applied to non-Gaussian projection and fitting errors.
  • domain assumption Novel categories share geometric shape with one of the hand-defined super categories built from base classes.
    Super Category Condition maps motorcycle/bicycle, pedestrian/cone, etc.; if a novel class does not fit any super category, refinement transfers wrong geometry.
  • domain assumption VLM 2D detections and SAM masks provide reliable seeds for novel objects.
    The whole pipeline starts from 2D detection boxes; missed or misplaced 2D detections cannot be recovered by later modules.
  • domain assumption The experimental class split groups base and novel classes by geometric similarity, making geometric transfer possible.
    Open-vocabulary settings section follows Yin et al.; the 7.37% result depends on this favorable split rather than arbitrary novel classes.
  • domain assumption GPT-4-provided size priors approximate real-world dimensions.
    Table 6 priors are used to construct candidate boxes and thresholds; no justification beyond LLM query.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HQ-OV3D: A High Box Quality Open-World 3D Detection Framework based on Diffision Model." pith.science (2026). https://pith.science/paper/XM2BH53I

@misc{pith2026250810935,
  author       = {Pith},
  title        = {Pith review of: HQ-OV3D: A High Box Quality Open-World 3D Detection Framework based on Diffision Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XM2BH53I}},
  note         = {Machine review of arXiv:2508.10935}
}
read the original abstract

Traditional closed-set 3D detection frameworks fail to meet the demands of open-world applications like autonomous driving. Existing open-vocabulary 3D detection methods typically adopt a two-stage pipeline consisting of pseudo-label generation followed by semantic alignment. While vision-language models (VLMs) recently have dramatically improved the semantic accuracy of pseudo-labels, their geometric quality, particularly bounding box precision, remains commonly neglected. To address this issue, we propose a High Box Quality Open-Vocabulary 3D Detection (HQ-OV3D) framework, dedicated to generate and refine high-quality pseudo-labels for open-vocabulary classes. The framework comprises two key components: an Intra-Modality Cross-Validated (IMCV) Proposal Generator that utilizes cross-modality geometric consistency to generate high-quality initial 3D proposals, and an Annotated-Class Assisted (ACA) Denoiser that progressively refines 3D proposals by leveraging geometric priors from annotated categories through a DDIM-based denoising mechanism. Compared to the state-of-the-art method, training with pseudo-labels generated by our approach achieves a 7.37% improvement in mAP on novel classes, demonstrating the superior quality of the pseudo-labels produced by our framework. HQ-OV3D can serve not only as a strong standalone open-vocabulary 3D detector but also as a plug-in high-quality pseudo-label generator for existing open-vocabulary detection or annotation pipelines.

Figures

Figures reproduced from arXiv: 2508.10935 by the authors.

Figure 1
Figure 1. Comparison between existing OV-3D pipeline and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the limitations of heuristic-based [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed HQ-OV3D framework. The pipeline consists of two main modules: the Intra-Modality Cross-Validated Proposal Generator, which generates proposals for novel classes, and the Annotated-Class Assisted Denoiser, which refines noisy proposals through a diffusion-based mechanism. Together, these modules enable high-quality pseudo-label generation for OV-3D. where x ′ i and y ′ i are the normalized of… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: This figure compares the novel-class pseudo-labels generated by our HQ-OV3D framework and Find n’ Propagate. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Detection results of the downstream detector [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 38 canonical work pages

  1. [1]

    Bai, X.; Hu, Z.; Zhu, X.; Huang, Q.; Chen, Y.; Fu, H.; and Tai, C.-L. 2022. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1090--1099

    work page 2022

  2. [2]

    H.; Vora, S.; Liong, V

    Caesar, H.; Bankiti, V.; Lang, A. H.; Vora, S.; Liong, V. E.; Xu, Q.; Krishnan, A.; Pan, Y.; Baldan, G.; and Beijbom, O. 2020. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11621--11631

    2020

  3. [3]

    Cao, Y.; Yihan, Z.; Xu, H.; and Xu, D. 2023. Coda: Collaborative novel box discovery and cross-modal alignment for open-vocabulary 3d object detection. Advances in Neural Information Processing Systems, 36: 71862--71873

    work page 2023

  4. [4]

    Chen, X.; Zhang, T.; Wang, Y.; Wang, Y.; and Zhao, H. 2023. Futr3d: A unified sensor fusion framework for 3d detection. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 172--181

    work page 2023

  5. [5]

    Cheng, T.; Song, L.; Ge, Y.; Liu, W.; Wang, X.; and Shan, Y. 2024. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 16901--16911

    work page 2024

  6. [6]

    Chow, A.; Riddell, E.; Wang, Y.; Sedwards, S.; and Czarnecki, K. 2025. OV-SCAN: Semantically Consistent Alignment for Novel Object Discovery in Open-Vocabulary 3D Object Detection. arXiv preprint arXiv:2503.06435

    work page Pith review arXiv 2025

  7. [7]

    Deng, J.; Shi, S.; Li, P.; Zhou, W.; Zhang, Y.; and Li, H. 2021. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, volume 35, 1201--1209

    work page 2021

  8. [8]

    Etchegaray, D.; Huang, Z.; Harada, T.; and Luo, Y. 2024. Find n’propagate: Open-vocabulary 3d object detection in urban environments. In European Conference on Computer Vision, 133--151. Springer

    work page 2024

Show all 47 references
  1. [9]

    Fan, L.; Xiong, X.; Wang, F.; Wang, N.; and Zhang, Z. 2021. Rangedet: In defense of range view for lidar-based 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, 2918--2927

    2021

  2. [10]

    C.; Lo, W.-Y.; et al

    Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A. C.; Lo, W.-Y.; et al. 2023. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, 4015--4026

    2023

  3. [11]

    H.; Vora, S.; Caesar, H.; Zhou, L.; Yang, J.; and Beijbom, O

    Lang, A. H.; Vora, S.; Caesar, H.; Zhou, L.; Yang, J.; and Beijbom, O. 2019. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12697--12705

    2019

  4. [12]

    H.; Zhang, P.; Zhang, H.; Yang, J.; Li, C.; Zhong, Y.; Wang, L.; Yuan, L.; Zhang, L.; Hwang, J.-N.; et al

    Li, L. H.; Zhang, P.; Zhang, H.; Yang, J.; Li, C.; Zhong, Y.; Wang, L.; Yuan, L.; Zhang, L.; Hwang, J.-N.; et al. 2022 a . Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10965--10975

    2022

  5. [13]

    Li, Y.; Chen, Y.; Qi, X.; Li, Z.; Sun, J.; and Jia, J. 2022 b . Unifying voxel-based representation with transformer for 3d object detection. Advances in Neural Information Processing Systems, 35: 18442--18455

    2022

  6. [14]

    Li, Z.; Wang, F.; and Wang, N. 2021. Lidar r-cnn: An efficient and universal 3d object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7546--7555

    2021

  7. [15]

    Li, Z.; Wang, W.; Li, H.; Xie, E.; Sima, C.; Lu, T.; Yu, Q.; and Dai, J. 2024. Bevformer: learning bird's-eye-view representation from lidar-camera via spatiotemporal transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2024

  8. [16]

    Liang, Z.; Zhang, M.; Zhang, Z.; Zhao, X.; and Pu, S. 2020. Rangercnn: Towards fast and accurate 3d object detection with range image representation. arXiv preprint arXiv:2009.00206

    2020 arXiv

  9. [17]

    Liang, Z.; Zhang, Z.; Zhang, M.; Zhao, X.; and Pu, S. 2021. Rangeioudet: Range image based real-time 3d object detector optimized by intersection over union. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7140--7149

    2021

  10. [18]

    Liu, S.; Zeng, Z.; Ren, T.; Li, F.; Zhang, H.; Yang, J.; Jiang, Q.; Li, C.; Yang, J.; Su, H.; et al. 2024. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European conference on computer vision, 38--55. Springer

    2024

  11. [19]

    Liu, Z.; Tang, H.; Lin, Y.; and Han, S. 2019. Point-voxel cnn for efficient 3d deep learning. Advances in neural information processing systems, 32

    2019

  12. [20]

    Lu, Y.; Xu, C.; Wei, X.; Xie, X.; Tomizuka, M.; Keutzer, K.; and Zhang, S. 2023. Open-vocabulary point-cloud object detection without 3d annotation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1190--1199

    2023

  13. [21]

    Luo, Y.; Chen, Z.; Fang, Z.; Zhang, Z.; Baktashmotlagh, M.; and Huang, Z. 2023 a . Kecor: Kernel coding rate maximization for active 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, 18279--18290

    2023

  14. [22]

    Luo, Y.; Chen, Z.; Wang, Z.; Yu, X.; Huang, Z.; and Baktashmotlagh, M. 2023 b . Exploring active 3d object detection from a generalization perspective. arXiv preprint arXiv:2301.09249

    2023 arXiv

  15. [23]

    Mao, J.; Xue, Y.; Niu, M.; Bai, H.; Feng, J.; Liang, X.; Xu, H.; and Xu, C. 2021. Voxel transformer for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, 3164--3173

    2021

  16. [24]

    P.; Laddha, A.; Kee, E.; Vallespi-Gonzalez, C.; and Wellington, C

    Meyer, G. P.; Laddha, A.; Kee, E.; Vallespi-Gonzalez, C.; and Wellington, C. K. 2019. Lasernet: An efficient probabilistic 3d object detector for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12677--12686

    2019

  17. [25]

    E.; and Huang, G

    Pan, X.; Xia, Z.; Song, S.; Li, L. E.; and Huang, G. 2021. 3d object detection with pointformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7463--7472

    2021

  18. [26]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learning, 8748--8763. PmLR

    2021

  19. [27]

    Sheng, H.; Cai, S.; Liu, Y.; Deng, B.; Huang, J.; Hua, X.-S.; and Zhao, M.-J. 2021. Improving 3d object detection with channel-wise transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 2743--2752

    2021

  20. [28]

    Shi, S.; Guo, C.; Jiang, L.; Wang, Z.; Shi, J.; Wang, X.; and Li, H. 2020. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10529--10538

    2020

  21. [29]

    Shi, S.; Jiang, L.; Deng, J.; Wang, Z.; Guo, C.; Shi, J.; Wang, X.; and Li, H. 2023. PV-RCNN++: Point-voxel feature set abstraction with local vector representation for 3D object detection. International Journal of Computer Vision, 131(2): 531--551

    2023

  22. [30]

    Shi, S.; Wang, X.; and Li, H. 2019. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 770--779

    2019

  23. [31]

    Shi, W.; and Rajkumar, R. 2020. Point-gnn: Graph neural network for 3d object detection in a point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1711--1719

    2020

  24. [32]

    Song, J.; Meng, C.; and Ermon, S. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502

    2020 arXiv

  25. [33]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  26. [34]

    C.; Zhang, T.; Wang, Y.; Zhao, H.; and Solomon, J

    Wang, Y.; Guizilini, V. C.; Zhang, T.; Wang, Y.; Zhao, H.; and Solomon, J. 2022. Detr3d: 3d object detection from multi-view images via 3d-to-2d queries. In Conference on robot learning, 180--191. PMLR

    2022

  27. [35]

    Yan, Y.; Mao, Y.; and Li, B. 2018. Second: Sparsely embedded convolutional detection. Sensors, 18(10): 3337

    2018

  28. [36]

    Yang, B.; Luo, W.; and Urtasun, R. 2018. Pixor: Real-time 3d object detection from point clouds. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 7652--7660

    2018

  29. [37]

    Yang, T.; Ju, Y.; and Yi, L. 2024. ImOV3D: Learning Open Vocabulary Point Clouds 3D Object Detection from Only 2D Images. Advances in Neural Information Processing Systems, 37: 141261--141291

    2024

  30. [38]

    Yang, Z.; Sun, Y.; Liu, S.; and Jia, J. 2020. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11040--11048

    2020

  31. [39]

    Yang, Z.; Sun, Y.; Liu, S.; Shen, X.; and Jia, J. 2018. Ipod: Intensive point-based object detector for point cloud. arXiv preprint arXiv:1812.05276

    2018 arXiv

  32. [40]

    Yang, Z.; Sun, Y.; Liu, S.; Shen, X.; and Jia, J. 2019. Std: Sparse-to-dense 3d object detector for point cloud. In Proceedings of the IEEE/CVF international conference on computer vision, 1951--1960

    2019

  33. [41]

    Yin, T.; Zhou, X.; and Krahenbuhl, P. 2021. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11784--11793

    2021

  34. [42]

    Zhang, H.; Wang, Y.; Dayoub, F.; and Sunderhauf, N. 2021. Varifocalnet: An iou-aware dense object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 8514--8523

    2021

  35. [43]

    Zhang, H.; Xu, J.; Tang, T.; Sun, H.; Yu, X.; Huang, Z.; and Yu, K. 2024. OpenSight: A simple open-vocabulary framework for LiDAR-based object detection. In European Conference on Computer Vision, 1--19. Springer

    2024

  36. [44]

    H.; and Koltun, V

    Zhao, H.; Jiang, L.; Jia, J.; Torr, P. H.; and Koltun, V. 2021. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 16259--16268

    2021

  37. [45]

    Zhou, Y.; and Tuzel, O. 2018. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, 4490--4499

    2018

  38. [46]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  39. [47]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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