Pith.

Computer Vision and Pattern Recognition · arXiv:2608.26656 v1 · submitted 27 Aug 2026 · 24 references

CoGeo-GS: Concept-Driven and Geometry-Aware Multi-Object Removal in 3D Scenes

Yuanxiang Ni, Xianliang Huang, Chenhang Ma, Chen Xiao, Yuewen Ma, Ruxin Wang, Hao Zhang

The paper says

“Multi-object removal in 3D scenes is challenging due to severe occlusions, semantic entanglement, and the difficulty of maintaining geometric and multi-view consistency.”

more of the abstract

Existing 3D Gaussian Splatting (3DGS) methods perform well for single-object editing but scale poorly to multi-object scenarios, often requiring repetitive optimization and yielding unstable geometry in removed regions. We propose CoGeo-GS, a concept-driven framework for controllable multi-object removal in 3D scenes. CoGeo-GS assigns concept-aware semantic tags to Gaussians, enabling flexible object selection and reducing interference between foreground objects and background structures within a single optimization stage. To recover plausible geometry, we introduce a geometry-aware completion pipeline that combines monocular depth priors with diffusion-based refinement and boundary-aligned blending. A geometry-regularized refinement strategy further stabilizes reconstruction and preserves multi-view consistency. Experiments demonstrate that CoGeo-GS outperforms existing methods in visual quality and reconstruction fidelity.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 9 Sep 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. CoGeo-GS makes a clear step forward on multi-object removal in 3DGS, but the lack of error bars makes the headline numbers hard to trust without more runs. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 24 references say.

The paper says, at Table I (Task 1 & Task 2)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

All quantitative results are reported as single-run values with no error bars, standard deviations, or significance tests. The central claim of state-of-the-art performance rests entirely on these point estimates. Given the stochasticity in 3DGS training, diffusion sampling, and baseline hyperparameters, it is essential to provide at least three independent runs (e.g., with different random seeds) and report mean ± std for each metric.

more of objection 1

The ablation study in Table II has the same limitation. Without this information, the reader cannot assess whether the reported improvements are reproducible or within the noise of the evaluation protocol.

The paper says, at Eq. (8) and §III-C (Stage 1)

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The scale alignment in Eq. (8) fits monocular depth to rendered depth using a least-squares criterion over background regions. When the background region is small (e.g., large occlusions or scenes with multiple large objects), the fit may be ill-conditioned, producing unstable alignment that propagates to the diffusion refinement stage.

more of objection 2

The paper does not analyze this failure mode or quantify the minimum required background pixel fraction. The authors should either provide an analysis (synthetic or real) of alignment stability under varying background coverage or discuss this limitation explicitly.

The paper says, at §IV-A (Experimental Setup)

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The manuscript states that baselines are retrained using official open-source implementations but does not specify whether identical random seeds, learning rates, or iteration counts were used across methods. To isolate the effect of the proposed approach from training-configuration differences, this information should be reported. Additionally, the number of views used for evaluation on each scene should be stated.

5 minor objections

The paper says, at Eq. (4)

Pith says: minor objection 1 of 5

The notation 'Ec∼O(u)' is non-standard; it appears to mean the expectation over the class distribution at pixel u. Please rewrite as 'E_{c∼O(u)}' or use the standard cross-entropy form '∑_{c} [c=O(u)] log p_c(u)'.

The paper says, at Fig. 1

Pith says: minor objection 2 of 5

The figure is information-dense. Consider labeling subfigures (a)–(d) more clearly in the caption and using distinct colors for different data flows (e.g., depth, RGB, identity). The text in subfigure (d) is small and hard to read.

show 3 more

The paper says, at §III-D (Gaussian Fusion)

Pith says: minor objection 3 of 5

The pruning thresholds δ_near, r, and n_min are listed in the implementation details but their values are not provided (e.g., δ_near = 0.05 world units?). Please specify the numerical values or explain how they are set.

The paper says, at §IV-A (Implementation Details)

Pith says: minor objection 4 of 5

The feature dimension D=16 and nearest neighbors k=5 are given, but the spatial consistency weight λ=0.1, trust region threshold ε, and SSIM weight λ_ssim=0.2 are mentioned only for the refinement stage. Please list all hyperparameter values in a single table for reproducibility.

The paper says, at References

Pith says: minor objection 5 of 5

Reference [4] is a closely related work by the same first author. While self-citation is not problematic, the connection between [4] and the current paper should be discussed in the text (it is only cited in the Related Work).

What would settle it

Run the multi-object removal benchmark (Table I, Task 1) three times with different random seeds for CoGeo-GS and the strongest baseline (Gaussian Grouping). Report mean ± std for all metrics. If the improvement remains at least 2 dB PSNR and the standard deviations do not exceed 0.5 dB, the claim is robust; otherwise, the advantage may be within noise.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 9 Sep 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 25 s · 9,838 tokens read, 2,536 written
  3. Editor wrote the verdict above 19 s · 10,618 tokens read, 1,899 written
  4. Referee major revision: 3 major objections, 5 minor 24 s · 10,097 tokens read, 2,317 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper tackles a practical and under-addressed problem: removing multiple objects from a 3D Gaussian scene in a single pass, with plausible geometry and appearance. The reported numbers in Table I are striking—3.3 dB PSNR gain over the strongest baseline, 56.6% LPIPS reduction—and the method is sensibly designed. The concept-aware tagging (Grounding DINO → SAM → distilled 3D features) is a clean way to enable multi-object selection without per-object optimization. The depth completion pipeline, which aligns monocular priors via least-squares on background then refines with a diffusion model under a scale-anchored constraint, is a reasonable integration. The trust-region clipping for geometric updates during fine-tuning is a neat engineering trick that plausibly stabilizes fusion boundaries.

The main soft spot is the lack of any error bars or multiple-run statistics. Every number in the paper comes from a single run. Given that 3DGS training involves randomness (SfM seeds, optimization stochasticity), and the baselines were retrained, we have no way to assess whether the reported gains are reproducible or within the noise of a lucky seed. The stress-test note is right to call this out. The improvements are large enough that they probably hold, but the claim would be much stronger with means and standard deviations over several seeds.

A secondary issue: the pipeline chains several pre-trained models (Grounding DINO, SAM, Depth Anything 3, InFusion). The paper does not discuss failure modes or how errors propagate. For example, a poor SAM mask in one view could corrupt the 3D labels and hurt subsequent steps. A brief failure analysis or robustness experiment would help.

Some hyperparameters (k=5, λ=0.1, trust-region ε, δ_near, r, n_min) are listed but not justified or tested for sensitivity. Minor, but easy to address.

Overall, this is a solid, well-built system with clear engineering contributions. It deserves peer review. A good referee will ask for multi-run statistics and a failure discussion. I’d bring it to a reading group focused on 3D scene editing, and I’d cite it if I worked on multi-object removal, though with a note about the evaluation limitations.

The strongest objection

The central claim that CoGeo-GS outperforms baselines rests solely on single-run metric values in Table I. Without error bars or multiple trials, it is impossible to assess whether the large reported gaps (e.g., 3.3 dB PSNR improvement, 56.6% LPIPS reduction) are reproducible or within the noise of random initialization, training stochasticity, and baseline hyperparameter sensitivity. This is particularly concerning because the paper retrains baselines but does not specify whether the same random seeds, learning rates, or number of iterations were used. The ablation study also reports single runs. The claim would be much more reliable with uncertainty quantification.

The reader's weakest assumption focused on depth alignment, which is a technical vulnerability but not necessarily fatal to the presented comparisons. However, the most load-bearing concern for the central claim (outperformance on benchmarks) is the lack of statistical rigor. The reader also noted 'lack of error bars on metrics' as a secondary concern in their rationale. My stress-test elevates this to the primary concern because without replication uncertainty, the numerical improvements could be artifacts of single trials. The proposed concrete test directly addresses this. Since the reader already flagged this and issued a CONDITIONAL verdict, no change is needed; the test would either validate or weaken the claim.

The referee's report

If the reported gains are reproducible, CoGeo-GS represents a meaningful step forward for multi-object removal in 3DGS, a task that prior methods handled only for single objects or with unstable geometry. The concept-driven tagging and the hybrid depth anchoring+diffusion completion are technically sound and address real problems (semantic entanglement, scale drift). However, the lack of error bars or multiple-run statistics across all quantitative tables prevents a firm assessment of statistical significance, making it unclear whether the large metric gaps (e.g., +3.3 dB PSNR, -56.6% LPIPS) reflect systematic improvement or noise.

Pith's reading of the paper (a machine's summary, not the authors')

The paper introduces CoGeo-GS, a framework for removing multiple objects from 3D scenes reconstructed with Gaussian Splatting. Existing methods handle single-object editing well but struggle with multi-object scenarios: they require repeated per-object optimization, entangle foreground and background, and produce unstable geometry in removed regions. CoGeo-GS assigns a learnable concept tag to each Gaussian, enabling simultaneous selection and removal of any object subset in one optimization stage while reducing interference between objects and background. To recover plausible geometry where objects are removed, it anchors monocular depth priors (Depth Anything 3) to the scene's coordinate system via least-squares alignment, then refines the depth with a diffusion-based inpainting model constrained to the masked region. A boundary-aware blending ensures smooth transitions, and a two-stage refinement (first freeze geometry, then allow small trust-region updates) stabilizes appearance and preserves multi-view consistency. Experiments on standard benchmarks (Mip-NeRF 360, SPIn-NeRF) show that CoGeo-GS outperforms GaussianEditor, InFusion, Gaussian Grouping, and SPIn-NeRF in both visual quality and reconstruction fidelity, with a 3.3 dB PSNR improvement for multi-object removal and reductions of 56.6% in LPIPS and 50.0% in FID.

What carries the argument. Concept-aware semantic tagging: each Gaussian learns a $16$-dimensional feature vector distilled from 2D object masks (obtained via text-driven detection and segmentation) using a multi-class cross-entropy loss and a 3D nearest-neighbor KL-divergence regularizer. Depth-guided geometric completion: a two-stage pipeline where (1) monocular depth (Depth Anything 3) is scale-aligned to rendered depth over background pixels by least-squares fitting (Eq.~8), and (2) a pre-trained depth diffusion model (InFusion) refines the depth within the hole while enforcing a scale-anchored constraint (Eq.~11). Trust-region refinement: after freezing geometry for appearance adaptation, geometric updates are clipped by norm (Eq.~13) to allow only small local corrections at fusion boundaries.

Core claim. The paper claims that by distilling 2D semantic masks into per-Gaussian concept features through a combination of cross-entropy loss and 3D spatial consistency regularization, the framework can identify and remove multiple objects in a single forward pass of optimization, without per-object repetition. When combined with a geometry completion pipeline that first aligns monocular depth predictions to the scene's global scale via background least-squares fitting and then refines hole geometry with a diffusion model under a masked constraint, the method produces geometrically consistent and visually coherent removals. The geometric regularization in the final refinement stage (freeze then trust-region updates) prevents structural drift while eliminating fusion seams. This unified approach is the first to achieve state-of-the-art multi-object removal within the 3D Gaussian Splatting paradigm.

The premise it rests on. The method assumes that monocular depth predictions (from Depth Anything 3) can be reliably aligned to the 3DGS geometry by a simple least-squares fit over background regions (Eq. 8); in scenes where the background is severely occluded or too small, this alignment becomes unstable, potentially causing the depth-guided completion to produce geometrically incorrect surfaces.

If the paper is right (Pith's inferences, not the authors')

Users can select arbitrary combinations of objects for removal by specifying text prompts, and the entire editing is completed in a single optimization stage, avoiding the computational cost of per-object pipelines.

The removed regions are filled with geometrically consistent depth and appearance, eliminating the black holes and floating artifacts typical of prior methods.

The concept tagging mechanism provides an explicit object-level representation that can be reused for other tasks such as object insertion, replacement, or compositional scene editing.

Quantitative metrics (PSNR, SSIM, LPIPS, FID) improve substantially over the strongest baseline, with the largest gains in perceptual quality (LPIPS reduced by 56.6%) for multi-object scenes.

The geometry-regularized refinement ensures that the final scene supports multi-view consistent rendering, critical for VR, robotics, and autonomous driving applications.

If the concept-tag distillation generalizes to unseen objects without retraining, the method could form the basis of an interactive 3D editor where users select objects by natural language and remove them in real time.

The depth anchor step relies on sufficient background pixels for least-squares fitting; in heavily occluded scenes (e.g., a cluttered desk with few visible background regions), the alignment may become unstable, which could be tested by designing a benchmark with occlusion ratios above 80%.

The diffusion-based depth refinement inherits biases from its training data; scenes with unusual geometry (e.g., thin structures, reflective surfaces) may produce implausible completions, pointing to a need for domain-adaptive depth priors.

The framework currently handles static scenes; extending to dynamic or time-varying environments would require mechanisms to update concept tags and geometry over time, which the paper does not address.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Multi-object removal in 3D scenes is challenging due to severe occlusions, semantic entanglement, and the difficulty of maintaining geometric and multi-view consistency. Existing 3D Gaussian Splatting (3DGS) methods perform well for single-object editing but scale poorly to multi-object scenarios, often requiring repetitive optimization and yielding unstable geometry in removed regions. We propose CoGeo-GS, a concept-driven framework for controllable multi-object removal in 3D scenes. CoGeo-GS assigns concept-aware semantic tags to Gaussians, enabling flexible object selection and reducing interference between foreground objects and background structures within a single optimization stage. To recover plausible geometry, we introduce a geometry-aware completion pipeline that combines monocular depth priors with diffusion-based refinement and boundary-aligned blending. A geometry-regularized refinement strategy further stabilizes reconstruction and preserves multi-view consistency. Experiments demonstrate that CoGeo-GS outperforms existing methods in visual quality and reconstruction fidelity.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23

The authors' main claim, as the classifier picked it out 1.00

“Experiments demonstrate that CoGeo-GS outperforms existing methods in visual quality and reconstruction fidelity.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.08 · human participants 0.03 · controlled comparison 0.60 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.93 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Multi-object removal in 3D scenes is challenging due to severe occlusions, semantic entanglement, and the difficulty of maintaining geometric and multi-view consistency. background 1.00 0.07 0.02 asserted 1.00
Existing 3D Gaussian Splatting (3DGS) methods perform well for single-object editing but scale poorly to multi-object scenarios, often requiring repetitive optimization and yielding unstable geometry in removed regions. gap 0.62 0.05 0.03 asserted 1.00
We propose CoGeo-GS, a concept-driven framework for controllable multi-object removal in 3D scenes. contribution 1.00 0.84 0.12 asserted 0.99
CoGeo-GS assigns concept-aware semantic tags to Gaussians, enabling flexible object selection and reducing interference between foreground objects and background structures within a single optimization stage. method 0.92 0.88 0.07 asserted 1.00
To recover plausible geometry, we introduce a geometry-aware completion pipeline that combines monocular depth priors with diffusion-based refinement and boundary-aligned blending. method 0.93 0.91 0.04 asserted 1.00
A geometry-regularized refinement strategy further stabilizes reconstruction and preserves multi-view consistency. method 0.88 0.81 0.06 asserted 1.00
Experiments demonstrate that CoGeo-GS outperforms existing methods in visual quality and reconstruction fidelity. result 1.00 0.95 0.84 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,552 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (4)
Figure 1 of the paper
Figure 1. Overview of the CoGeo-GS framework. (a) We first distill text-driven segmentation masks into 3D Gaussians to assign object-level identities. (b)
Figure 2 of the paper
Figure 2. Qualitative comparison on multi-object removal. We compare CoGeo-GS with GaussianEditor, InFusion and Gaussian Grouping in a consistent view.
Figure 3 of the paper
Figure 3. Qualitative comparison on single-object removal. We compare CoGeo
Figure 4 of the paper
Figure 4. Ablation study. Without concept-aware masking, residual artifacts

Figures are the authors', reproduced from the paper.

References (24)
  1. [1] Haotian Mao, Zhuoxiong Xu, Siyue Wei, Yule Quan, Nianchen Deng, and Xubo Yang, “Live-gs: Llm powers interactive vr by enhancing gaussian splatting,”arXiv preprint arXiv:2412.09176, 2024 on PithPith's reviewarXiv
  2. [2] Zhenjie Yang, Yilin Chai, Xiaosong Jia, Qifeng Li, Yuqian Shao, Xuekai Zhu, Haisheng Su, and Junchi Yan, “Drivemoe: Mixture-of-experts for vision-language-action model in end-to-end autonomous driving,”arXiv preprint arXiv:2505.16278, 2025 on PithPith's reviewarXiv
  3. [3] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang, “3d gaussian splatting in robotics: A survey,”arXiv preprint arXiv:2410.12262, 2024 on PithPith's reviewarXiv
  4. [4] Xianliang Huang, Chen Xiao, Yuanxiang Ni, Guanming Liu, Mingkai Liu, Dikai Fan, Xiao Liu, and Hao Zhang, “Semantic-guided pro- gressive object removal with gaussian splatting,”arXiv preprint arXiv:2607.04144, 2026 on PithPith's reviewarXiv
  5. [5] Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis, “3d gaussian splatting for real-time radiance field rendering.,” ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023 on Pith
  6. [6] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin, “Gaussianeditor: Swift and controllable 3d editing with gaussian splatting,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2024, pp. 21476–21485 on Pith
  7. [7] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke, “Gaussian grouping: Segment and edit anything in 3d scenes,” inProc. Eur. Conf. Comput. Vis.Springer, 2024, pp. 162–179 on Pith
  8. [8] Yuxin Wang, Qianyi Wu, Guofeng Zhang, and Dan Xu, “Learning 3d geometry and feature consistent gaussian splatting for object removal,” inProc. Eur. Conf. Comput. Vis.Springer, 2024, pp. 1–17 on Pith
  9. [9] Antonio Criminisi, Patrick P ´erez, and Kentaro Toyama, “Region filling and object removal by exemplar-based image inpainting,”IEEE Trans. Image Process., vol. 13, no. 9, pp. 1200–1212, 2004 on Pith
  10. [10] Tijana Ru ˇzi´c and Aleksandra Pi ˇzurica, “Context-aware patch-based image inpainting using markov random field modeling,”IEEE Trans. Image Process., vol. 24, no. 1, pp. 444–456, 2014 on Pith
  11. [11] Yu Zeng, Zhe Lin, Huchuan Lu, and Vishal M Patel, “Cr-fill: Generative image inpainting with auxiliary contextual reconstruction,” inProc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 14164–14173 on Pith
  12. [12] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky, “Resolution-robust large mask inpainting with fourier convolutions,” inProc. IEEE/CVF Winter Conf. Appl... on Pith
  13. [13] Jonathan Ho, Ajay Jain, and Pieter Abbeel, “Denoising diffusion probabilistic models,”Adv. Neural Inf. Process. Syst., vol. 33, pp. 6840– 6851, 2020 on Pith
  14. [14] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole, “Score-based generative modeling through stochastic differential equations,”arXiv preprint arXiv:2011.13456, 2020 on PithPith's reviewarXiv
  15. [15] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool, “Repaint: Inpainting using denoising diffusion probabilistic models,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 11461–11471 on Pith
  16. [16] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang, “Smartbrush: Text and shape guided object inpainting with diffusion model,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2023, pp. 22428–22437 on Pith
  17. [17] Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Der- panis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein, “Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields,” inProc. IEEE/CVF Conf. Comput. Vis. P... on Pith
  18. [18] Xianliang Huang, Shuhang Chen, Zhizhou Zhong, Jiajie Gou, Jihong Guan, and Shuigeng Zhou, “Hi-nerf: Hybridizing 2d inpainting with neural radiance fields for 3d scene inpainting,” inProceedings of the Asian Conference on Computer Vision, 2024, pp. 2855–2871 on Pith
  19. [19] Xianliang Huang, Zhizhou Zhong, Shuhang Chen, Yi Xu, Jihong Guan, and Shuigeng Zhou, “Nerf-mir: Toward high-quality restoration of masked images with neural radiance fields,”IEEE Transactions on Neural Networks and Learning Systems, 2026 on Pith
  20. [20] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao, “Infusion: Inpainting 3d gaussians via learning depth completion from diffusion prior,”arXiv preprint arXiv:2404.11613, 2024 on PithPith's reviewarXiv
show 4 more
  1. [21] Kira Prabhu, Jane Wu, Lynn Tsai, Peter Hedman, Dan B Goldman, Ben Poole, and Michael Broxton, “Inpaint3d: 3d scene content generation using 2d inpainting diffusion,”arXiv preprint arXiv:2312.03869, 2023 on PithPith's reviewarXiv
  2. [22] Xianliang Huang, Jiajie Gou, Shuhang Chen, Zhizhou Zhong, Jihong Guan, and Shuigeng Zhou, “Iddr-ngp: Incorporating detectors for distractors removal with instant neural radiance field,” inProceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 1343–1351 on Pith
  3. [23] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman, “Mip-nerf 360: Unbounded anti-aliased neural radiance fields,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit., 2022, pp. 5470–5479 on Pith
  4. [24] Johannes L Schonberger and Jan-Michael Frahm, “Structure-from- motion revisited,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 4104–4113 on Pith
Cited by 1 Pith paper
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 3 sentences without a confident label

Compares with 1 paper
Cites as background 5 papers

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment