Pith. sign in

REVIEW 3 major objections 6 minor 63 references

Seg-Wild: Interactive Segmentation based on 3D Gaussian Splatting for Unconstrained Image Collections

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Seg-Wild claims that interactive 3D segmentation works on scenes reconstructed from unconstrained Internet photo collections, not just carefully captured images.

desk verdict Useful first step toward interactive segmentation of unconstrained 3DGS scenes, but the quantitative claim is not yet trustworthy because the evaluation shares SAM as both teacher and grader. read the letter →

arxiv 2507.07395 v1 pith:IJMMBTHI submitted 2025-07-10 cs.CV

classification cs.CV
keywords 3DGaussianSplattinginteractivesegmentationunconstrainedphotocollectionsSegmentAnythingModelin-the-wildscenereconstructionaffinityfeaturefieldscale-adaptivespikycutter
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

Seg-Wild claims that interactive 3D segmentation can be done on scenes reconstructed from unconstrained Internet photo collections — the kind with inconsistent lighting, changing weather, and tourists walking through the frame — rather than only on carefully captured images. It stores a learned feature vector on every 3D Gaussian, trains those features while reconstructing the scene, and lets a user click a target to select the Gaussians whose features match. To keep object boundaries clean, it uses SAM's masks with adaptive prompt density; to remove protruding splats, it trims spiky Gaussians according to how much of them the mask covers. If the claim holds, object selection, novel-view rendering, and appearance tuning become available for casual photo sets, which lowers the cost of collecting data for 3D editing.

What carries the argument

The load-bearing object is the per-Gaussian affinity feature $af_i$, a learned vector appended to each 3D Gaussian's parameters $(p_i,\Sigma_i,\alpha_i,sf_i,df_i,af_i)$. It is trained by minimizing $L_{FE}=\|fe_{SAM}-fe_{rend}\|_1$, the $\ell^1$ distance between SAM's PCA-compressed 64-dimensional features and the splat-rendered features, together with a compactness loss $L_{Com}$ that weights rendered-feature cosines by SAM mask IoU so Gaussians inside the same mask move closer together. The scale-adaptive module (SASM) chooses SAM's grid scale from camera-to-Gaussian distance and distributes prompt points by a sky-filtered depth map; the Spiky 3D Gaussian Cutter (SGC) projects each spiky Gaussian's covariance, computes how much of its long axis the SAM mask covers, and shrinks the Gaussian by that ratio. Together these pieces convert an unconstrained photo collection into a selectable 3D scene.

What would settle it

Re-run Seg-Wild on the Photo Tourism benchmark with SAM masks deliberately corrupted in a known way, such as erasing the upper half of a target in every view; if the final 3D segmentation still contains the erased region, the method is not gated by SAM masks, while if the output reproduces the erased boundary, the claimed robustness to unconstrained photos is bounded by SAM's mask quality.

Watch

Extended reading notes

Core claim

The paper's central claim is that lifting SAM's 2D features into a 3D affinity feature field inside 3D Gaussian Splatting makes interactive segmentation of in-the-wild scenes work, and that the combination of a SAM-mask compactness loss, a depth-adaptive SAM prompt scale, and a spiky-Gaussian cutter outperforms Feature 3DGS, SAGA, and GS-W with projection-based segmentation on both IoU and Accuracy for every target in the Photo Tourism benchmark. The same optimization also reconstructs the scene, so segmentation inherits the appearance and transient-occlusion handling of the GS-W reconstruction it builds on. The authors additionally introduce a segmentation benchmark on the Photo Tourism and NeRF-On-the-go datasets, with three targets per scene.

Load-bearing premise

The load-bearing premise is that SAM's 2D masks are accurate and consistent enough across the whole photo collection, because those masks steer the compactness loss, gate which Gaussians enter the final segmentation, and decide how much of each spiky Gaussian is cut.

Editorial extensions

If this is right

  • A user can select an object in a reconstructed landmark from ordinary tourist photos with a single click, requiring no 3D annotations.
  • Selected regions can be rendered from new viewpoints and their lighting can be tuned by interpolating appearance weights, enabling downstream editing of the segmented object.
  • Data collection for 3D segmentation reduces to whatever Internet photos exist of a site, because the method handles transient occluders and lighting changes.
  • The new benchmark gives later methods a shared set of in-the-wild targets, reported as IoU and Accuracy on the Photo Tourism scenes.

Reading between the lines

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

  • Editorial inference: the method inherits SAM's 2D boundary errors; the reported failure case where the winged horse's wings are missing suggests that improvements in 2D foundation models would transfer directly into segmentation gains here.
  • Editorial inference: the affinity-feature machinery is not tied to SAM specifically, so text-conditioned or open-vocabulary features could plausibly replace the mask-gated pipeline and turn clicks into language prompts.
  • Editorial inference: the benchmark covers only three scenes and three targets each, so the quantitative claim is a first demonstration rather than a proven guarantee across all unconstrained collections.
  • Editorial inference: the SGC ratio-trimming rule is a generic post-process for any Gaussian-based segmentation with a 2D mask, and could be applied independently of the wild-scene training pipeline.
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

3 major / 6 minor

Summary. The paper proposes Seg-Wild, an interactive 3D segmentation method for scenes reconstructed from unconstrained photo collections with 3D Gaussian Splatting (3DGS). It adds per-Gaussian affinity features trained with a SAM-based compactness loss, a scale-adaptive SAM module (SASM) that adjusts the segmentation scale and sampling points based on projected Gaussian depth, and a Spiky Gaussian Cutter (SGC) that post-processes the initial feature-similarity segmentation to reduce protruding Gaussians. The reconstruction backbone is GS-W. Quantitative evaluation is reported on three Photo Tourism scenes with nine target objects, comparing against Feature 3DGS, SAGA, and GS-W with projection-based segmentation; supplementary material includes qualitative results, including a failure case and appearance tuning. The abstract and conclusion additionally claim improved reconstruction quality.

Significance. Interactive segmentation of 3D scenes reconstructed from casual Internet photo collections is a valuable and underexplored task. The paper is the first, to my knowledge, to combine SAM-based 3D feature fields with an in-the-wild reconstruction backbone (GS-W) and to provide a segmentation benchmark for Photo Tourism scenes. The ablation study (Table 2) is a genuine strength: each proposed component (L_Com, SASM, sky-mask filtering, SGC, feature dimensionality) is ablated and shows a measurable contribution, and the failure-case discussion in the supplementary is honest. However, the headline quantitative claim is not yet established because of the evaluation confounds detailed below. If those are addressed, the work could be a practical step forward for downstream tasks on Internet photo collections.

major comments (3)
  1. [§4.1, Eq. (7)] The compactness loss L_Com is written as a double sum over all H×W pixel pairs. For the Photo Tourism images used here (roughly 800×600 or larger), this requires on the order of 10^11–10^12 pair evaluations per image per iteration, which is computationally intractable. The paper does not specify any sampling strategy, number of pairs, or mask subsampling. Because Table 2, row (1), shows that removing L_Com causes a large performance drop, the actual computation of this loss is load-bearing. Please specify exactly how L_Com is computed in practice, including the number of sampled pixel pairs and how the SAM masks are used to form S.
  2. [§4.3 (Eqs. 19, 21) and §B.1] The evaluation protocol is confounded by the use of the same SAM mask on both sides of the benchmark. Eq. (19) admits a Gaussian only when its projected center is inside the SAM mask M2D, and Eq. (21) uses the same M2D to decide how much of each Gaussian to cut. Separately, Section B.1 states that the ground-truth masks in the benchmark are initialized from SAM masks and then manually refined. The reported IoU/Acc in Table 1 are therefore bounded from above by the quality of the SAM mask M2D, and the method cannot recover object parts omitted by M2D—exactly the failure documented in Fig. 11. This makes the Table 1 caption claim of "superior results across all metrics and targets" unverifiable as an assessment of the 3D feature field, SASM, or SGC. I recommend adding (i) an ablation of Seg-Wild without the M2D gate, (ii) a reference row reporting the IoU/Acc of M2D alone, and (iii) baselines (Feature 3DGS and SAGA) with the same M2D gating applied.
  3. [§2.1, §5.2, Table 1, abstract] The central claims exceed the experimental evidence in three related ways. First, WildSeg3D [15], which the paper itself describes as a real-time interactive segmentation method designed for in-the-wild scenes, is not included in Table 1. Second, the abstract and conclusion claim improved "reconstruction quality," but no PSNR, SSIM, LPIPS, or any other reconstruction metric is reported anywhere in the paper or the supplementary material. Third, Section 5.2 states that the method is evaluated on both Photo Tourism and NeRF-On-the-go, but the only quantitative table uses three PT scenes, and the NeRF-On-the-go results are limited to qualitative novel-view synthesis (Fig. 9). Please add the missing baseline and metrics, or revise the claims to match what is actually measured.
minor comments (6)
  1. [§A.4, Eq. (28)] The sentence "s_max and s_min are set to 4 and 8, respectively, representing the upper and lower limits" is internally inconsistent; the range is stated as [4, 8] in B.2 and the main text. The assignment of s_max and s_min should be corrected to match that range.
  2. [§4.2, Eqs. (14)–(15)] N_PP(i,j) is described as the "number of prompt points assigned to that grid," but Eq. (15) generates N_PP × N_PP points. Please clarify that N_PP is the number of points per row and column, or change the wording to "number of prompt points per side."
  3. [§C.1] The sentence "As shown in the second column of Figure 8, although SAGA performs feature compaction" should refer to the third column, since SAGA is the third compared method in that figure and the second column is Feature 3DGS, as in the main-text Figure 4.
  4. [§4.1, Eq. (6)] The notation <·,·> is declared to denote cosine similarity, but the same operation is written as a normalized inner product in Eq. (16). Please use one consistent notation throughout.
  5. [§2.1] There are typos: "performd" and "deliverd" should be "performed" and "delivered."
  6. [§5.1] The implementation details state only the training time. Please report the input image resolution used for training and for SAM mask generation, as the SASM sampling cost and the L_Com pair count both depend on it.

Circularity Check

1 steps flagged · score 2.0 of 10

Headline superiority is partly an evaluation loop: Eq. 19 hard-gates Seg-Wild's output by the SAM mask while the benchmark ground truth (Sec. B.1) starts from SAM prompt-masks, so reported IoU/Acc gains are not fully independent; the core method is otherwise empirical and not derived.

  1. self definitional [Sec. 4.3, Eq. (19)-(22); benchmark construction in Sec. B.1 (supplement)]
    "Finally use the prompt points to generate a SAM mask M2D ... obtaining the segmentation result M3D: ... M3D = {p_j | s_fus > tau, M2D(u_j,v_j)=1}. ... For benchmark generation, we employed the open-source ISAT framework in conjunction with the Segment Anything Model (SAM). Specifically, we used SAM's built-in prompt-based segmentation to generate initial 2D masks. These masks were manually refined to improve ground truth quality."

    The final 3D segmentation M3D is defined to contain only Gaussians whose 2D projection lies inside the SAM mask M2D (Eq. 19), and the SGC post-processor uses the same M2D to decide how much of each spiky Gaussian to cut (Eqs. 21-22). The ground-truth masks used to compute the Table 1 IoU/Acc are initialized from SAM prompt-based segmentation and then manually refined (Sec. B.1). Thus the reported scores largely measure agreement with SAM's own mask rather than an independent 3D object boundary: any Gaussian outside M2D is excluded by construction, and the benchmark annotation shares the same source. Manual refinement softens but does not break the loop; the Fig. 11 failure case (SAM omits the wings) is the same coupling in reverse.

full rationale

Seg-Wild makes no first-principles derivation claim; its contributions are algorithmic modules validated empirically on Photo Tourism and NeRF-On-the-go. I found no load-bearing self-citation (GS-W is an external prior, and the coauthor-cited works are not used to justify the core segmentation), no imported uniqueness theorem, no ansatz smuggled by citation, and no renamed known result. The one concrete circular signal is in the evaluation design: the final 3D mask is hard-gated by the SAM 2D mask (Eq. 19), the SGC post-processor uses the same SAM mask for its coverage ratio (Eqs. 21-22), and the benchmark ground truth is initialized from SAM's prompt-based masks then manually refined (Sec. B.1). Consequently, the headline 'superior results across all metrics and targets' is partly a measure of how closely the method reproduces SAM's own mask, not purely an independent 3D segmentation quality. Because the ground truth is manually refined, because the compared baselines also use SAM, and because reconstruction quality and ablations provide separate evidence, this is an evaluation-design confounding rather than a derivation that reduces to its inputs. The scale range [4,8] and threshold tau=0.5 are selected on the same benchmark scenes, which is a further external-validity caveat but not a circular derivation. Overall, the central method retains independent empirical content, so the circularity is minor (score 2).

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

The method's novelty rests on empirically tuned modules (SASM, SGC) and on inheriting GS-W and SAM assumptions. Hyperparameters like the [4,8] scale range and tau are hand-chosen, and several are tuned on the benchmark used for evaluation.

free parameters (5)
  • Segmentation scale range [s_min, s_max] = 4 to 8
    Chosen as the "optimal normalization range derived from our experiments" (Supplementary B.2); tuned on the same benchmark used for evaluation, a potential source of overfitting.
  • Similarity threshold tau = 0.5
    Threshold for including Gaussians in M3D (Eq. 19); ablation shows 0.3 and 0.7 perform slightly worse, so 0.5 was selected by hand.
  • Affinity feature channel count C = 64
    Feature dimension after PCA compression; ablation (Table 2 rows 7-9) shows 128 similar or slightly better but chosen as a trade-off for compute.
  • Loss weights lambda_FE and lambda_Com = 0.7 and 0.3
    Set in Eq. 8 without ablation; presumably tuned informally.
  • Sky mask confidence threshold = 0.5
    Set in Supplementary B.3 to preserve details; not ablated across scenes.
assumptions (5)
  • domain assumption GS-W reconstruction supplies accurate geometry and appearance decomposition for unconstrained photo collections.
    Seg-Wild builds directly on GS-W and inherits its assumptions about transient occlusion masks and appearance embeddings (Sections 3 and A.2).
  • domain assumption SAM feature embeddings are stable and discriminative across views with varying lighting and occlusions.
    The affinity feature field is trained to reproduce SAM features (Eqs. 4-5); if SAM features are inconsistent across views, the 3D features will be noisy.
  • ad hoc to paper The depth-based sampling heuristic (Eq. 14) improves SAM mask granularity for unconstrained scenes.
    No theoretical justification; average depth is treated as a proxy for semantic density, with hard caps of 1 and 20 points per grid cell.
  • ad hoc to paper The linear depth-to-scale mapping (Eq. 28) with range [4,8] is appropriate for all scenes.
    Range was derived from experiments on the evaluation scenes (Supplementary B.2), not from a general principle, and is not tested on held-out scenes.
  • domain assumption Sky regions should be suppressed when distributing SAM prompt points.
    Sky depth is replaced with the minimum depth (Eq. 12) so foreground receives more prompts; this fails if the sky itself is the segmentation target.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seg-Wild: Interactive Segmentation based on 3D Gaussian Splatting for Unconstrained Image Collections." pith.science (2026). https://pith.science/paper/IJMMBTHI

@misc{pith2026250707395,
  author       = {Pith},
  title        = {Pith review of: Seg-Wild: Interactive Segmentation based on 3D Gaussian Splatting for Unconstrained Image Collections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJMMBTHI}},
  note         = {Machine review of arXiv:2507.07395}
}
read the original abstract

Reconstructing and segmenting scenes from unconstrained photo collections obtained from the Internet is a novel but challenging task. Unconstrained photo collections are easier to get than well-captured photo collections. These unconstrained images suffer from inconsistent lighting and transient occlusions, which makes segmentation challenging. Previous segmentation methods cannot address transient occlusions or accurately restore the scene's lighting conditions. Therefore, we propose Seg-Wild, an interactive segmentation method based on 3D Gaussian Splatting for unconstrained image collections, suitable for in-the-wild scenes. We integrate multi-dimensional feature embeddings for each 3D Gaussian and calculate the feature similarity between the feature embeddings and the segmentation target to achieve interactive segmentation in the 3D scene. Additionally, we introduce the Spiky 3D Gaussian Cutter (SGC) to smooth abnormal 3D Gaussians. We project the 3D Gaussians onto a 2D plane and calculate the ratio of 3D Gaussians that need to be cut using the SAM mask. We also designed a benchmark to evaluate segmentation quality in in-the-wild scenes. Experimental results demonstrate that compared to previous methods, Seg-Wild achieves better segmentation results and reconstruction quality. Our code will be available at https://github.com/Sugar0725/Seg-Wild.

Figures

Figures reproduced from arXiv: 2507.07395 by the authors.

Figure 1
Figure 1. (a) Our method uses unconstrained photo collec [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of our framework. During the reconstruction of in-the-wild scenes, we embed affinity features into [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The procedure for SASM is to generate prompt [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: We perform a qualitative comparison of segmentation results across four methods: Feature 3DGS [ [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visualization of segmentation results before and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Visualization of sky masks under different confi [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of segmentation results [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 11
Figure 11. Figure 11: Visualization of Failure Cases. The performance [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 9
Figure 9. Figure 9: Visualization results of the novel view synthesis [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Visualization results of appearance tuning on the [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 44 canonical work pages

  1. [15]

    Yansong Guo, Jie Hu, Yansong Qu, and Liujuan Cao. 2025. WildSeg3D: Segment Any 3D Objects in the Wild from 2D Images. arXiv preprint arXiv:2503.08407 (2025)

  2. [1]

    Francisco Massa Adam Lerer James Bradbury Gregory Chanan Trevor Killeen Zeming Lin Natalia Gimelshein Luca Antiga Alban Desmaison Andreas Köpf Edward Yang Zach DeVito Martin Raison Alykhan Tejani Sasank Chilamkurthy Benoit Steiner Lu Fang Junjie Bai Adam Paszke, Sam Gross and Soumith Chintala

  3. [2]

    Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. 2011. Building rome in a day. Commun. ACM 54, 10 (2011), 105–112

  4. [3]

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. 2021. Mip-nerf: A multiscale repre- sentation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision . 5855–5864

  5. [4]

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision. 9650–9660

  6. [5]

    Jiazhong Cen, Jiemin Fang, Chen Yang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. 2025. Segment any 3d gaussians. In Proceedings of the AAAI Confer- ence on Artificial Intelligence. 1971–1979

  7. [6]

    Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al . 2023. Segment anything in 3d with nerfs. In Advances in Neural Information Processing Systems . 25971–25990

  8. [7]

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. 2022. Tensorf: Tensorial radiance fields. In European conference on computer vision . 333–350

Show all 63 references
  1. [8]

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. 2018. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision . 801–818

  2. [9]

    Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do

  3. [10]

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus En- zweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. 2016. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and patt...

  4. [11]

    Bin Dou, Tianyu Zhang, Zhaohui Wang, Yongjia Ma, Zejian Yuan, and Nanning Zheng. 2025. Learning segmented 3D Gaussians via efficient feature unprojection for zero-shot neural scene segmentation. In International Conference on Neural Information Processing. 398–412

  5. [12]

    Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. 2020. Points2surf learning implicit surfaces from point clouds. In Proceedings of the European Conference on Computer Vision . 108–124

  6. [13]

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. 2023. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12479–12488

  7. [14]

    Rahul Goel, Dhawal Sirikonda, Saurabh Saini, and PJ Narayanan. 2023. Interactive segmentation of radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 4201–4211

  8. [16]

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)

  9. [17]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  10. [18]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

  11. [19]

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al

  12. [20]

    Jonas Kulhanek, Songyou Peng, Zuzana Kukelova, Marc Pollefeys, and Torsten Sattler. 2024. Wildgaussians: 3d gaussian splatting in the wild. arXiv preprint arXiv:2407.08447 (2024)

  13. [21]

    Loic Landrieu and Martin Simonovsky. 2018. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition . 4558–4567

  14. [22]

    Boyi Li, Kilian Q Weinberger, Serge Belongie, Vladlen Koltun, and Rene Ranftl

  15. [23]

    In Proceedings of the IEEE/CVF international conference on computer vision

    Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision. 4015–4026

  16. [24]

    Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. 2021. Nerf in the wild: Neural radiance fields for unconstrained photo collections. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

  17. [25]

    Moustafa Meshry, Dan B Goldman, Sameh Khamis, Hugues Hoppe, Rohit Pandey, Noah Snavely, and Ricardo Martin-Brualla. 2019. Neural rerendering in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6878–6887

  18. [26]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis. Commun. ACM 65, 1 (2021), 99–106

  19. [27]

    Ashkan Mirzaei, Tristan Aumentado-Armstrong, Konstantinos G Derpanis, Jonathan Kelly, Marcus A Brubaker, Igor Gilitschenski, and Alex Levinshtein

  20. [28]

    Xueting Li, Sifei Liu, Shalini De Mello, Kihwan Kim, Xiaolong Wang, Ming-Hsuan Yang, and Jan Kautz. 2020. Online adaptation for consistent mesh reconstruction in the wild. In Advances in Neural Information Processing Systems . 15009–15019

  21. [29]

    Charles Ruizhongtai Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. 2017. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition . 652–660

  22. [30]

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. 2017. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems . 5105–5114

  23. [31]

    Yansong Qu, Dian Chen, Xinyang Li, Xiaofan Li, Shengchuan Zhang, Liujuan Cao, and Rongrong Ji. 2025. Drag Your Gaussian: Effective Drag-Based Editing with Score Distillation for 3D Gaussian Splatting. arXiv preprint arXiv:2501.18672 (2025)

  24. [32]

    Yansong Qu, Shaohui Dai, Xinyang Li, Jianghang Lin, Liujuan Cao, Shengchuan Zhang, and Rongrong Ji. 2024. Goi: Find 3d gaussians of interest with an opti- mizable open-vocabulary semantic-space hyperplane. In Proceedings of the 32nd ACM International Conference on Multimedia ....

  25. [33]

    In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Spin-nerf: Multiview segmentation and perceptual inpainting with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20669–20679

  26. [34]

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El- Nouby, et al. 2023. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

  27. [35]

    Weining Ren, Zihan Zhu, Boyang Sun, Jiaqi Chen, Marc Pollefeys, and Songyou Peng. 2024. Nerf on-the-go: Exploiting uncertainty for distractor-free nerfs in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 8931–8940

  28. [36]

    Zhongzheng Ren, Aseem Agarwala, Bryan Russell, Alexander G Schwing, and Oliver Wang. 2022. Neural volumetric object selection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 6133–6142

  29. [37]

    Sara Sabour, Lily Goli, George Kopanas, Mark Matthews, Dmitry Lagun, Leonidas Guibas, Alec Jacobson, David Fleet, and Andrea Tagliasacchi. 2025. SpotLessS- plats: Ignoring Distractors in 3D Gaussian Splatting. ACM Trans. Graph. 44, 2 (2025), 1–11

  30. [38]

    You Shen, Zhipeng Zhang, Xinyang Li, Yansong Qu, Yu Lin, Shengchuan Zhang, and Liujuan Cao. 2025. Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization. arXiv preprint arXiv:2503.00881 (2025)

  31. [39]

    Yansong Qu, Yuze Wang, and Yue Qi. 2023. SG-NeRF: Semantic-guided Point- based Neural Radiance Fields. In2023 IEEE International Conference on Multimedia and Expo. 570–575

  32. [40]

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. 2024. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

  33. [41]

    Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snavely. 2022. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings . 1–9

  34. [42]

    Yuze Wang, Junyi Wang, Ruicheng Gao, Yansong Qu, Wantong Duan, Shuo Yang, and Yue Qi. 2025. Look at the Sky: Sky-aware Efficient 3D Gaussian Splatting in the Wild. IEEE Transactions on Visualization and Computer Graphics 31, 5 (2025), 3481–3491

  35. [43]

    Yuze Wang, Junyi Wang, and Yue Qi. 2024. WE-GS: An In-the-wild Efficient 3D Gaussian Representation for Unconstrained Photo Collections. arXiv preprint arXiv:2406.02407 (2024). Seg-Wild, 2025, 07 Bao et al

  36. [44]

    Yuze Wang, Junyi Wang, Yansong Qu, and Yue Qi. 2023. Rip-nerf: Learning rotation-invariant point-based neural radiance field for fine-grained editing and compositing. In Proceedings of the 2023 ACM international conference on multimedia retrieval. 125–134

  37. [45]

    Noah Snavely, Steven M Seitz, and Richard Szeliski. 2006. Photo tourism: ex- ploring photo collections in 3D. In ACM SIGGRAPH 2006 Conference Proceedings . 835–846

  38. [46]

    Shuang Song, Zhaopeng Cui, and Rongjun Qin. 2021. Vis2mesh: Efficient mesh reconstruction from unstructured point clouds of large scenes with learned virtual view visibility. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 6514–6524

  39. [47]

    Jiacong Xu, Yiqun Mei, and Vishal Patel. 2024. Wild-gs: Real-time novel view syn- thesis from unconstrained photo collections. In Advances in Neural Information Processing Systems. 103334–103355

  40. [48]

    Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. 2024. Gaussian grouping: Segment and edit anything in 3d scenes. InProceedings of the European Conference on Computer Vision. 162–179

  41. [49]

    Qiao Yu, Xianzhi Li, Yuan Tang, Jinfeng Xu, Long Hu, Yixue Hao, and Min Chen

  42. [50]

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. 2024. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 19447–19456

  43. [51]

    Yuze Wang, Junyi Wang, Chen Wang, Wantong Duan, Yongtang Bao, and Yue Qi. 2024. SCARF: Scalable Continual Learning Framework for Memory-efficient Multiple Neural Radiance Fields. In Computer Graphics Forum , Vol. 43. Wiley Online Library, e15255

  44. [52]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing 13, 4 (2004), 600–612

  45. [53]

    Boming Zhao, Bangbang Yang, Zhenyang Li, Zuoyue Li, Guofeng Zhang, Jiashu Zhao, Dawei Yin, Zhaopeng Cui, and Hujun Bao. 2022. Factorized and controllable neural re-rendering of outdoor scene for photo extrapolation. In Proceedings of the 30th ACM international conference on mu...

  46. [54]

    Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. 2024. Feature 3dgs: Supercharging 3d gaussian splatting to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Co...

  47. [55]

    BG”, “TM

    Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. 2023. Segment everything everywhere all at once. In Advances in Neural Information Processing Systems . 19769–19782. Seg-Wild: Interactive Segmentation based on...

  48. [56]

    arXiv preprint arXiv:2406.15811 (2024)

    PointDreamer: Zero-shot 3D Textured Mesh Reconstruction from Colored Point Cloud by 2D Inpainting. arXiv preprint arXiv:2406.15811 (2024)

  49. [58]

    Dongbin Zhang, Chuming Wang, Weitao Wang, Peihao Li, Minghan Qin, and Haoqian Wang. 2024. Gaussian in the wild: 3d gaussian splatting for uncon- strained image collections. In European Conference on Computer Vision . Springer, 341–359

  50. [59]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  51. [2018]

    In Proceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition . 586–595

  52. [2019]

    In Advances in Neural Information Processing Systems

    Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems . 8026–8037

  53. [2022]

    In International Conference on Learning Representations

    Language-driven Semantic Segmentation. In International Conference on Learning Representations

  54. [2023]

    ACM Trans

    3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph. 42, 4 (2023), 139:1–139:14

  55. [2024]

    InProceedings of the European Conference on Computer Vision

    Click-gaussian: Interactive segmentation to any 3d gaussians. InProceedings of the European Conference on Computer Vision . 289–305

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.