Pith. sign in

REVIEW 3 major objections 4 minor 138 references

Ground-truth boxes, shown to a teacher during training, improve visual grounding beyond serving as coordinate labels.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 04:14 UTC pith:OSFT67S7

load-bearing objection A modest but real empirical contribution: the privileged-box claim survives the ablations, but the paper needs a wrong-box control, multi-seed runs, and code before the exact size of the effect is fully trustworthy. the 3 major comments →

arxiv 2607.15732 v2 pith:OSFT67S7 submitted 2026-07-17 cs.CV

IoU-PD: IoU-Aware Privileged Distillation for Visual Grounding with Multimodal Large Language Models

Xiuyuan Zhu , Ke Lu , Hao Wu , Siwen Jiao , Zijin Du , Dongming Zhang , Jian Xue This is my paper
classification cs.CV
keywords visual groundingprivileged distillationcoordinate generationmultimodal large language modelIoU-aware weightingreferring expressionknowledge distillationon-policy self-distillation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Visual grounding with multimodal large language models is usually trained to output bounding-box coordinates as text, but scored by geometric overlap between predicted and true boxes. IoU-PD uses ground-truth boxes twice: as the supervised coordinate target, and as a privileged visual hint drawn on the image for a frozen teacher model. The teacher sees the box-marked image plus a short hint, while the student sees only the original image and prompt; the student is trained with a supervised fine-tuning anchor and an IoU-aware distillation loss. At inference the teacher, hint, and box overlay are all removed, so the method changes only training, not the deployed interface. Across standard referring-expression benchmarks, the full setting improves mIoU by 4.03 points and Acc@0.7 by 4.25 points over a strong 4B base model, with ablations attributing the gain to the box-marked teacher and the IoU-aware weighting.

Core claim

The paper claims that ground-truth boxes carry useful training-time guidance beyond their role as output labels. When a frozen teacher conditions on the image with the target box drawn in green and a text hint saying the answer lies inside the green rectangle, its output distribution contains region-aware information that a student, seeing only the original image, can absorb through distillation. The student is anchored by supervised fine-tuning on the true coordinate string and shaped by a token-weighted KL loss whose weights combine a sample-level IoU factor, per-coordinate error, digit-position significance, teacher–student agreement, and teacher confidence. Experiments show consistent re

What carries the argument

The central mechanism is a frozen privileged teacher paired with an on-policy distillation loss. The teacher receives the original image with the ground-truth box overlaid in green, plus the original prompt augmented with the hint 'The answer is located within the green rectangle'; the student receives only the original image and prompt. The student's own decoded box is scored under both branches, and the KL term is token-weighted by a product of factors — exponential in (1 − IoU of the sampled box), softmax over coordinate-level errors, digit-position significance, a sigmoid on teacher–student log-probability agreement, and teacher entropy — so that geometrically important and reliable toke

Load-bearing premise

The load-bearing premise is that the frozen teacher's distribution, computed on an input that literally contains the ground-truth box, remains a faithful distillation target for a student that will never see that box—that the conditional mismatch between the privileged and standard inputs is small enough that the teacher's logits teach region awareness rather than reliance on the green rectangle.

What would settle it

Retrain the same pipeline with a teacher whose green rectangle is placed at a random location (same size, same hint text); if the student's mIoU and Acc@0.7 gains stay close to the full IoU-PD numbers, the claim that ground-truth location information drives the improvement is falsified, whereas a drop to the original-teacher baseline would confirm that the privileged location is the active component.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Training-time-only privileged visual hints can be layered onto coordinate-generating multimodal LLMs without changing the inference input format or adding a localization head.
  • The gains increase under stricter overlap thresholds: P@0.9 rises 7.75 points and P@0.95 by 12.84 over the base model, indicating the method pushes many predictions into the high-overlap range.
  • The method helps small and medium objects more than large ones, suggesting privileged box hints are especially useful when coordinate errors cover a larger fraction of the target region.
  • Removing the SFT anchor sharply reduces the gain, so privileged distillation complements direct coordinate supervision rather than replacing it.
  • The approach requires grounding annotations at training time; without ground-truth boxes, the privileged teacher input cannot be constructed.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the mechanism would be to replace the ground-truth overlay with a randomly placed green box of the same size; if the gain persists, the active ingredient is the visual marker itself rather than the location information.
  • The same privileged-distillation scheme could transfer to other structured-output tasks with geometric supervision and spatial metrics, such as keypoint detection, dense captioning, or layout generation.
  • Because the teacher is frozen and sees the base distribution plus the overlay, the method is a form of on-policy self-distillation with a visual hint; it may combine naturally with verifiable rewards that directly optimize IoU.
  • The paper's own limitation—that the hint is a simple overlay and gains are moderate—points to richer privileged inputs, such as a cropped region or segmentation mask, as a way to reduce the remaining conditional mismatch.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes IoU-PD, a training-time privileged-distillation method for coordinate-generating multimodal large language models. The student is trained on the original image and referring expression, while a frozen teacher receives a box-marked image plus an appended hint. The total loss is an SFT anchor plus a token-weighted KL distillation over student-generated responses, with weights informed by IoU, coordinate error, digit position, teacher-student agreement, and teacher entropy. At inference the teacher, box overlay, and hint are removed. Experiments on RefCOCO, RefCOCO+, and RefCOCOg with Qwen3-VL-4B report consistent gains over the base model (+4.03 mIoU, +4.25 Acc@0.7 in the main setting), with ablations isolating SFT, original-teacher distillation, box-marked teacher input, and IoU-aware weighting. The paper honestly lists limitations, including the requirement for grounding annotations and the condition-mismatch risk of privileged distillation.

Significance. If the reported gains are robust, the paper makes a practical and conceptually useful contribution: it improves region-level grounding without changing the inference-time interface, adding a teacher branch, or requiring extra inference-time annotations. The ablation structure is a genuine strength: the SFT anchor, the non-privileged self-distillation baseline, and the IoU-aware weighting are separately tested, and the threshold, object-size, and scaling analyses give a more complete picture than a single benchmark number would. The central claim—that ground-truth boxes can serve as privileged guidance beyond coordinate labels—is plausible and largely well supported by the component ablations. However, the key isolation of the box-teacher effect is confounded by the simultaneously introduced text hint, and every result is single-run with no error bars. These issues are fixable within the manuscript's scope but need to be addressed before the central causal claim can be accepted.

major comments (3)
  1. [§4.3, Table 7, Appendix E] The comparison that isolates the privileged box teacher is Variant B ('Original teacher') vs Variant D ('Box teacher'). Moving from B to D changes two variables at once: the green box overlay and the appended hint text 'The answer is located within the green rectangle.' There is no box-only, hint-only, wrong-box, or random-box control. The observed difference (+0.65 Acc@0.7, +0.0051 mIoU) is therefore not uniquely attributable to the semantic content of the ground-truth box. Since this is the central claim of the paper, please add a wrong-box teacher (same overlay and hint but with a non-target box) and a box-overlay-without-hint variant; if a wrong-box teacher produces a similar gain, the effect would be consistent with a peaked/regularizing target rather than with privileged visual guidance about the referred region.
  2. [§4.3, Table 7; §4.2, Table 3] All results are reported from single training runs with no error bars or significance tests. The load-bearing component differences are small: B-to-D is 0.65 Acc@0.7, D-to-E is 0.27, E-to-F is 0.08, and F-to-Full is 0.13. These differences are within the range of typical run-to-run variation for LLM fine-tuning, so the finer-grained claims (e.g., that the sample-level IoU factor with tau_r=0.5 helps over tau_r=1.0) are not reliably supported. Please provide multiple seeds with standard deviations, or an equivalent statistical analysis, for at least the Table 7 ablation rows and the main Table 6 comparison.
  3. [§3.3, Eq. (2); Appendix C] The method's mechanism requires that the frozen teacher's distribution p_T(·|x_t,·), which is conditioned on the box-marked image, be a learnable and beneficial target for the student's p_θ(·|x_s,·), which does not see the box. Appendix C acknowledges the conditional-mismatch risk but offers only a qualitative argument. No quantitative measurement is provided for how large the mismatch is, how it evolves during training, or whether the KL term functions as genuine guidance rather than a regularizer. Measuring teacher-student KL divergence on held-out prompts before and after training, or reporting teacher accuracy under x_t and student accuracy under x_s, would directly address the load-bearing assumption and materially strengthen the paper.
minor comments (4)
  1. [Table 2, row 'Original teacher'] Variant B improves mIoU and Acc@0.7 over SFT-only (0.8492 vs 0.8470; 85.63 vs 85.32) but degrades Acc@0.5 (89.80 vs 90.62). The paper does not discuss this cross-metric inconsistency; a brief explanation would help.
  2. [§3.4, Eq. (3)] The normalization expression for w̄_t is ambiguous as typeset ('|Y|w_t P_j...'). It should read w̄_t = |Y| w_t / (Σ_j w_j + ε); please fix the notation.
  3. [Table 3 and Figure 4] The 2B scaling point is only reported at 30k/1 epoch, so the comparison between backbone sizes is incomplete. Adding a 2B/300k/3-epoch run (or explicitly stating why it is omitted) would make the scaling conclusion cleaner.
  4. [Algorithm 1, line 5] The notation 'ŷ ← Decode(p_θ(·|x_s))' is underspecified. Clarify whether this is greedy decoding, sampling, or beam search, and whether the same decoding is used for the teacher's condition.

Circularity Check

0 steps flagged

Empirical training method; gains measured on held-out splits — no circular reduction. Score 2 reflects one minor non-load-bearing self-citation (Smooth Operator) and the acknowledged-but-unquantified teacher-student conditional mismatch.

full rationale

The central claim—that ground-truth boxes provide useful privileged guidance beyond coordinate labels—is an empirical effect, not a derivation that restates its inputs. The training objective (Section 3.3, L = L_kd + αL_sft) uses the GT box both as the SFT target and as the teacher's box overlay; the claimed gains (+4.03 mIoU, +4.25 Acc@0.7 over Qwen3-VL-4B) are measured on the held-out RefCOCO/RefCOCO+/RefCOCOg splits under a unified evaluation protocol (Table 6), so no fitted value is renamed as a prediction. The IoU-aware weights w_t = r(u)·c_t·d_t·a_t·e_t (Section 3.4) are per-sample training reweightings computed from the student's own decoded box; aligning the loss with the evaluation metric is metric-aligned optimization, not a by-construction guarantee of test-set improvement, and the ablations (Table 7, rows D->E->Full) show small, non-forced increments. The only load-bearing assumption—that the frozen box-marked teacher's distribution p_T(·|x_t) transfers to the student that lacks the box—is an acknowledged risk, not a circularity: the paper explicitly flags the conditional-mismatch danger in Appendix C ('increasing the conditional mismatch between teacher and student'), and the B-vs-D ablation is confounded (the hint text changes alongside the box) and lacks a wrong-box control, which is a validity/robustness limitation rather than an equivalence-by-construction. The paper's own Limitations section acknowledges that the method requires grounding annotations and that gains are moderate. The one self-citation is Smooth Operator (Jiao et al., 2026), a co-authored related-work baseline that is not load-bearing. No equation is equivalent to its inputs by definition.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The method introduces six hand-chosen weighting hyperparameters and relies on the standard privileged-distillation assumption that teacher-side conditional distributions transfer to the student. No new physical or architectural entities are postulated; the box overlay is an input transform, not an invented entity. The free-parameter count is moderate and the lack of sensitivity analysis beyond tau_r is a weakness.

free parameters (6)
  • SFT coefficient alpha = 1.0
    Chosen by hand to balance distillation and supervised fine-tuning; no sensitivity analysis reported (Appendix A.1).
  • Sample-level IoU temperature tau_r = 0.5
    Ablated on test splits (Variant F vs Full) and the better value selected; tau_r=1.0 gives a nearly identical result (+4.12 vs +4.25 Acc@0.7).
  • Coordinate softmax temperature tau_c = 1.0
    Chosen for the coordinate-error factor c_t; no sensitivity analysis reported (Appendix A.3).
  • Digit-position weight lambda_d = 0.5
    Scales the digit-position factor d_t; chosen by hand, no sensitivity analysis reported (Appendix A.3).
  • Agreement sigmoid slope beta = 3.0
    Controls the teacher-student agreement factor a_t; chosen by hand, no sensitivity analysis reported (Appendix A.3).
  • Teacher-confidence entropy temperature tau_e = 1.0
    Scales the entropy-based teacher confidence factor e_t; chosen by hand, no sensitivity analysis reported (Appendix A.3).
axioms (4)
  • domain assumption The teacher distribution p_T(.|x_t, y_hat_<t) computed under the box-marked privileged input is a useful distillation target for the student conditioned on the original input x_s.
    Section 3.3, Eq. (2); if the privileged input makes the teacher target too different from the student's inference-time input, distillation could mislead; the paper relies on this transfer.
  • domain assumption The token-to-coordinate parsing and the fallback rule that counts malformed responses as invalid are fair across all compared models.
    Appendix B.2; the unified evaluation protocol assumes the single parser does not systematically penalize any baseline.
  • domain assumption Digit-position significance (hundreds > tens > ones) and the multiplicative weighting factors reflect the true effect of token errors on IoU.
    Section 3.4; the weighting formula w_t = r(u).c_t.d_t.a_t.e_t assumes these heuristics align with geometric importance.
  • domain assumption The frozen same-architecture teacher initialized from the same base checkpoint provides reliable logits without EMA updates.
    Appendix A.2; no EMA, stop-gradient; assumes no catastrophic drift in teacher quality during training.

pith-pipeline@v1.3.0-alltime-deepseek · 17387 in / 15016 out tokens · 148675 ms · 2026-08-04T04:14:11.287755+00:00 · methodology

0 comments
read the original abstract

Visual grounding with multimodal large language models is commonly formulated as autoregressive coordinate generation, where a model outputs bounding-box coordinates as text given an image and a referring-expression prompt. While this interface is simple and compatible with instruction following, it introduces a mismatch between training and evaluation: training optimizes token-level likelihood over coordinate strings, whereas grounding quality is measured by geometric overlap. We propose IoU-PD, an IoU-aware privileged distillation method for coordinate-generating multimodal large language models. IoU-PD uses ground-truth boxes not only as coordinate targets, but also as privileged training-time guidance. During training, the student receives the original image and prompt, while a frozen teacher receives a box-marked image and an augmented prompt that indicates the marked region. The student is trained with a supervised fine-tuning anchor and a privileged distillation loss whose token weights reflect both geometric importance and teacher reliability. At inference time, IoU-PD requires no box overlay, privileged hint, teacher branch, or additional prediction module. Experiments on standard referring-expression grounding benchmarks show consistent region-level improvements over strong coordinate-generating baselines, demonstrating that ground-truth boxes can provide useful privileged guidance beyond serving as coordinate labels.

Figures

Figures reproduced from arXiv: 2607.15732 by Dongming Zhang, Hao Wu, Jian Xue, Ke Lu, Siwen Jiao, Xiuyuan Zhu, Zijin Du.

Figure 1
Figure 1. Figure 1: Overview of IOU-PD. Ground-truth boxes are used not only as coordinate targets, but also to construct privileged teacher inputs during training. The student receives the original image and original referring-expression prompt, while the teacher receives a box-marked image and an augmented prompt that indicates the marked region. The training objective combines an SFT anchor with IoU-aware privileged distil… view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of IoU-aware token weighting. tokens. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Component ablations of IOU-PD. The upper panel shows the enabled training components, and the lower panel reports Acc@0.5 and Acc@0.7 gains over the base model under the main setting. this setting, the teacher receives the same original image as the student. It improves Acc@.7 from 85.32 to 85.63 over SFT, but the gain is smaller than using a box-marked teacher. Replacing the original teacher with the priv… view at source ↗
Figure 4
Figure 4. Figure 4: Scaling ablations under different data sizes, epoch budgets, and final model settings. The curves show [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: IOU-PD improves grounding across ob￾ject sizes. Examples are grouped by the ground-truth bounding-box area in the normalized coordinate space: small (< 5%), medium (5%−10%), and large (> 10%). Points report P@0.5 and P@0.7 for the base model and the IOU-PD model, with orange segments and labels indicating absolute gains. only improving easy large-object cases, but also helps the model localize more size-se… view at source ↗
Figure 6
Figure 6. Figure 6: further examines how the improvement changes across IoU thresholds. Compared with the 4B base model, IOU-PD improves P@0.5 by 2.98 points, P@0.7 by 4.25 points, P@0.9 by 7.75 points, and P@0.95 by 12.84 points. The IoU dis￾tribution shows the same trend: predictions be￾low 0.5 IoU decrease from 11.4% to 8.4%, while predictions above 0.95 IoU increase from 35.7% to 48.5%. These results suggest that the priv… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of different privileged visual [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

138 extracted references · 34 linked inside Pith

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    arXiv preprint arXiv:2306.15195 , year=

    Shikra: Unleashing multimodal llm's referential dialogue magic , author=. arXiv preprint arXiv:2306.15195 , year=

    Pith/arXiv arXiv

  9. [9]

    International Conference on Learning Representations , year=

    Pix2seq: A Language Modeling Framework for Object Detection , author=. International Conference on Learning Representations , year=

  10. [10]

    arXiv preprint arXiv:2306.14824 , year=

    Kosmos-2: Grounding multimodal large language models to the world , author=. arXiv preprint arXiv:2306.14824 , year=

    Pith/arXiv arXiv

  11. [11]

    Advances in Neural Information Processing Systems , volume=

    Cogvlm: Visual expert for pretrained language models , author=. Advances in Neural Information Processing Systems , volume=

  12. [12]

    Ferret: Refer and Ground Anything Anywhere at Any Granularity , author=

  13. [13]

    European Conference on Computer Vision , pages=

    Griffon: Spelling out all object locations at any granularity with large language models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  14. [14]

    3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =

    Kerbl, Bernhard and Kopanas, Georgios and Leimk. 3D Gaussian Splatting for Real-Time Radiance Field Rendering , journal =

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Wu, Guanjun and Yi, Taoran and Fang, Jiemin and Xie, Lingxi and Zhang, Xiaopeng and Wei, Wei and Liu, Wenyu and Tian, Qi and Wang, Xinggang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  16. [16]

    European Conference on Computer Vision , pages=

    Citygaussian: Real-time high-quality large-scale scene rendering with gaussians , author=. European Conference on Computer Vision , pages=. 2025 , organization=

    2025

  17. [17]

    arXiv preprint arXiv:2404.14249 , year=

    CLIP-GS: CLIP-Informed Gaussian Splatting for Real-time and View-consistent 3D Semantic Understanding , author=. arXiv preprint arXiv:2404.14249 , year=

    Pith/arXiv arXiv

  18. [18]

    ECCV , year=

    CompGS: Smaller and Faster Gaussian Splatting with Vector Quantization , author=. ECCV , year=

  19. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  20. [20]

    European Conference on Computer Vision (ECCV) , year=

    Language-Driven Physics-Based Scene Synthesis and Editing via Feature Splatting , author=. European Conference on Computer Vision (ECCV) , year=

  21. [21]

    ECCV , year=

    Gaussian Grouping: Segment and Edit Anything in 3D Scenes , author=. ECCV , year=

  22. [22]

    2024 , booktitle =

    Fang, Jiemin and Wang, Junjie and Zhang, Xiaopeng and Xie, Lingxi and Tian, Qi , title =. 2024 , booktitle =

    2024

  23. [23]

    arXiv:2304.02643 , year=

    Segment Anything , author=. arXiv:2304.02643 , year=

    Pith/arXiv arXiv

  24. [24]

    2024 , eprint=

    SAM 2: Segment Anything in Images and Videos , author=. 2024 , eprint=

    2024

  25. [25]

    arXiv preprint arXiv:2303.05499 , year=

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection , author=. arXiv preprint arXiv:2303.05499 , year=

    Pith/arXiv arXiv

  26. [26]

    2024 , eprint=

    Grounded SAM: Assembling Open-World Models for Diverse Visual Tasks , author=. 2024 , eprint=

    2024

  27. [27]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Zhou, Hongyu and Shao, Jiahao and Xu, Lu and Bai, Dongfeng and Qiu, Weichao and Liu, Bingbing and Wang, Yue and Geiger, Andreas and Liao, Yiyi , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  28. [28]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Qin, Minghan and Li, Wanhua and Zhou, Jiawei and Wang, Haoqian and Pfister, Hanspeter , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  29. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Shi, Jin-Chuan and Wang, Miao and Duan, Hao-Bin and Guan, Shao-Hua , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  30. [30]

    International Conference on Computer Vision (ICCV) , year =

    Kerr, Justin* and Kim, Chung Min* and Goldberg, Ken and Kanazawa, Angjoo and Tancik, Matthew , title =. International Conference on Computer Vision (ICCV) , year =

  31. [31]

    Proceedings of the Winter Conference on Applications of Computer Vision (WACV) , month =

    Chacko, Rohan and Haeni, Nicolai and Khaliullin, Eldar and Sun, Lin and Lee, Douglas , title =. Proceedings of the Winter Conference on Applications of Computer Vision (WACV) , month =. 2025 , pages =

    2025

  32. [32]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    OpenGaussian: Towards Point-Level 3D Gaussian-based Open Vocabulary Understanding , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  33. [33]

    ECCV , year =

    Pixel-GS: Density Control with Pixel-aware Gradient for 3D Gaussian Splatting , author =. ECCV , year =

  34. [34]

    arXiv preprint arXiv:2411.19290 , year=

    SADG: Segment Any Dynamic Gaussian Without Object Trackers , author=. arXiv preprint arXiv:2411.19290 , year=

    arXiv

  35. [35]

    2023 , journal=

    Segment Any 3D Gaussians , author=. 2023 , journal=

    2023

  36. [36]

    2024 , eprint=

    Semantic Gaussians: Open-Vocabulary Scene Understanding with 3D Gaussian Splatting , author=. 2024 , eprint=

    2024

  37. [37]

    European Conference on Computer Vision , year=

    Mini-Splatting: Representing Scenes with a Constrained Number of Gaussians , author=. European Conference on Computer Vision , year=

  38. [38]

    2024 , journal =

    TRIPS: Trilinear Point Splatting for Real-Time Radiance Field Rendering , author=. 2024 , journal =

    2024

  39. [39]

    arXiv preprint arXiv:2403.19615 , year=

    SA-GS: Scale-Adaptive Gaussian Splatting for Training-Free Anti-Aliasing , author=. arXiv preprint arXiv:2403.19615 , year=

    Pith/arXiv arXiv

  40. [40]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Multi-scale 3d gaussian splatting for anti-aliased rendering , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  41. [41]

    2025 , volume =

    Qiu, Shi and Xie, Binzhu and Liu, Qixuan and Heng, Pheng-Ann , booktitle =. 2025 , volume =. doi:10.1109/AIxVR63409.2025.00039 , publisher =

    arXiv 2025

  42. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  43. [43]

    The Thirteenth International Conference on Learning Representations , year=

    OmniRe: Omni Urban Scene Reconstruction , author=. The Thirteenth International Conference on Learning Representations , year=

  44. [44]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Yang, Ziyi and Gao, Xinyu and Zhou, Wen and Jiao, Shaohui and Zhang, Yuqing and Jin, Xiaogang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  45. [45]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Huang, Yi-Hua and Sun, Yang-Tian and Yang, Ziyi and Lyu, Xiaoyang and Cao, Yan-Pei and Qi, Xiaojuan , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  46. [46]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  47. [47]

    and Millis, Bryan A

    Zhao, Mengyang and Liu, Quan and Jha, Aadarsh and Deng, Ruining and Yao, Tianyuan and Mahadevan-Jansen, Anita and Tyska, Matthew J. and Millis, Bryan A. and Huo, Yuankai , title =. 2021 , isbn =. doi:10.1007/978-3-030-87589-3_45 , booktitle =

    doi:10.1007/978-3-030-87589-3_45 2021

  48. [48]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Kerr, Justin and Kim, Chung Min and Goldberg, Ken and Kanazawa, Angjoo and Tancik, Matthew , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2023 , pages =

    2023

  49. [49]

    European Conference on Computer Vision (ECCV) , year =

    Anpei Chen and Zexiang Xu and Andreas Geiger and Jingyi Yu and Hao Su , title =. European Conference on Computer Vision (ECCV) , year =

  50. [50]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Weakly Supervised 3D Open-vocabulary Segmentation , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  51. [51]

    Proceedings of the 38th International Conference on Machine Learning , pages =

    Learning Transferable Visual Models From Natural Language Supervision , author =. Proceedings of the 38th International Conference on Machine Learning , pages =. 2021 , editor =

    2021

  52. [52]

    and Tancik, Matthew and Barron, Jonathan T

    Mildenhall, Ben and Srinivasan, Pratul P. and Tancik, Matthew and Barron, Jonathan T. and Ramamoorthi, Ravi and Ng, Ren , title =. 2021 , issue_date =. doi:10.1145/3503250 , journal =

    doi:10.1145/3503250 2021

  53. [53]

    ACM Trans

    Thomas M\""uller and Alex Evans and Christoph Schied and Alexander Keller , title =. ACM Trans. Graph. , issue_date =. 2022 , pages =. doi:10.1145/3528223.3530127 , publisher =

    arXiv 2022

  54. [54]

    and Kowalski, Marek and Johnson, Matthew and Shotton, Jamie and Valentin, Julien , title =

    Garbin, Stephan J. and Kowalski, Marek and Johnson, Matthew and Shotton, Jamie and Valentin, Julien , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2021 , pages =

    2021

  55. [55]

    2022 , booktitle=

    Plenoxels: Radiance Fields without Neural Networks , author=. 2022 , booktitle=

    2022

  56. [56]

    2025 , issn =

    DSEM-NeRF: Multimodal feature fusion and global–local attention for enhanced 3D scene reconstruction , journal =. 2025 , issn =. doi:https://doi.org/10.1016/j.inffus.2024.102752 , author =

    arXiv 2025

  57. [57]

    Pumarola, Albert and Corona, Enric and Pons-Moll, Gerard and Moreno-Noguer, Francesc , booktitle=

  58. [58]

    2024 , isbn =

    Lin, Youtian , title =. 2024 , isbn =. doi:10.1609/aaai.v38i4.28138 , booktitle =

    doi:10.1609/aaai.v38i4.28138 2024

  59. [59]

    ICCV , year=

    Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields , author=. ICCV , year=

  60. [60]

    Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Yu, Zehao and Chen, Anpei and Huang, Binbin and Sattler, Torsten and Geiger, Andreas , title =. Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  61. [61]

    CVPR , year=

    SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering , author=. CVPR , year=

  62. [62]

    arXiv preprint arXiv:2403.06908 , year=

    FreGS: 3D Gaussian Splatting with Progressive Frequency Regularization , author=. arXiv preprint arXiv:2403.06908 , year=

    Pith/arXiv arXiv

  63. [63]

    ECCV , year=

    Street Gaussians for Modeling Dynamic Urban Scenes , author=. ECCV , year=

  64. [64]

    CVPR , year =

    VastGaussian: Vast 3D Gaussians for Large Scene Reconstruction , author =. CVPR , year =

  65. [65]

    2025 , eprint=

    A Hierarchical Compression Technique for 3D Gaussian Splatting Compression , author=. 2025 , eprint=

    2025

  66. [66]

    arXiv preprint arXiv:2403.17898 , year=

    Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians , author=. arXiv preprint arXiv:2403.17898 , year=

    Pith/arXiv arXiv

  67. [67]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Matrixcity: A large-scale city dataset for city-scale neural rendering and beyond , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  68. [68]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Turki, Haithem and Ramanan, Deva and Satyanarayanan, Mahadev , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2022 , pages =

    2022

  69. [69]

    A density-based algorithm for discovering clusters in large spatial databases with noise , year =

    Ester, Martin and Kriegel, Hans-Peter and Sander, J\"". A density-based algorithm for discovering clusters in large spatial databases with noise , year =. Proceedings of the Second International Conference on Knowledge Discovery and Data Mining , pages =

  70. [70]

    doi:10.21105/joss.00205 , year =

    Leland McInnes and John Healy and Steve Astels , title =. doi:10.21105/joss.00205 , year =

    doi:10.21105/joss.00205

  71. [71]

    Xu and Jun-Mei Song and Mingchuan Zhang and Y

    Zhihong Shao and Peiyi Wang and Qihao Zhu and R. Xu and Jun-Mei Song and Mingchuan Zhang and Y. K. Li and Yu Wu and Daya Guo , booktitle =. ArXiv , title =

  72. [72]

    Nature , volume=

    Deepseek-r1 incentivizes reasoning in llms through reinforcement learning , author=. Nature , volume=. 2025 , publisher=

    2025

  73. [73]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Alvar, Saeed Ranjbar and Singh, Gursimran and Akbari, Mohammad and Zhang, Yong , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  74. [74]

    European conference on computer vision , pages=

    Mmbench: Is your multi-modal model an all-around player? , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  75. [75]

    2024 , eprint=

    MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models , author=. 2024 , eprint=

    2024

  76. [76]

    arXiv preprint arXiv:2506.10967 , year=

    Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs , author=. arXiv preprint arXiv:2506.10967 , year=

    Pith/arXiv arXiv

  77. [77]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year=

    FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age for Bias Measurement and Mitigation , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year=

  78. [78]

    arXiv preprint arXiv:2411.19930 , year=

    On Domain-Adaptive Post-Training for Multimodal Large Language Models , author=. arXiv preprint arXiv:2411.19930 , year=

    Pith/arXiv arXiv

  79. [79]

    arXiv preprint arXiv:2412.05271 , year=

    Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling , author=. arXiv preprint arXiv:2412.05271 , year=

    Pith/arXiv arXiv

  80. [80]

    arXiv preprint arXiv:2411.10442 , year=

    Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization , author=. arXiv preprint arXiv:2411.10442 , year=

    Pith/arXiv arXiv

Showing first 80 references.