Pith. sign in

REVIEW 5 major objections 5 minor 32 references

DivAS claims that interactive 3D segmentation of NeRF scenes can match per-scene optimized methods without any optimization, by depth-refining SAM masks and fusing them in a voxel grid.

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 ↗

An optimization-free pipeline lifts Segment Anything masks into NeRF scenes via depth-weighted refinement and CUDA voxel aggregation, matching optimization-based segmentation accuracy on LLFF and Mip-NeRF360 at 2–2.5× lower end-to-end cost.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection DivAS's core claim—training-free parity with optimized NeRF segmenters—is plausible but rests on one biased benchmark and a six-scene comparison with no error bars. the 5 major comments →

arxiv 2601.04860 v2 pith:MBESLMJA submitted 2026-01-08 cs.CV

DivAS: Interactive 3D Segmentation by Depth-Weighted Voxel Aggregation

Ayush Pande , Mayank Vatsa This is my paper
classification cs.CV
keywords interactive 3D segmentationNeural Radiance FieldsSegment Anything Modelzero-shot segmentationvoxel aggregationdepth-guided refinementCUDA kerneloptimization-free
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.

The reading

The paper aims to show that interactive 3D segmentation of a pre-trained NeRF does not need a representation-specific optimization loop. It proposes DivAS, which refines 2D foundation-model (SAM) masks with depth cues from the NeRF, then fuses the refined masks across a small set of anchor views into a probabilistic 3D voxel grid using a fast CUDA kernel. On the NVOS/LLFF benchmark, DivAS reports a higher mean IoU (0.909) than the optimization-based baselines SA3D (0.902) and SANeRF-HQ (0.887); on Mip-NeRF360 it reports a tied mIoU (0.913) with SANeRF-HQ while running end-to-end 2-2.5x faster. If correct, this would mean zero-shot 2D segmentation can be lifted to 3D at state-of-the-art quality without per-scene training, making interactive 3D segmentation practical with a few clicks.

Core claim

On its own terms, the paper's central claim is that a depth-weighted refinement of SAM masks, guided by the NeRF's rendered depth, followed by a voxel-aggregation kernel with geometric consistency checks, is sufficient to match the segmentation quality of per-scene optimization. The kernel projects every voxel into every centroid view, applies density, ray-voxel distance, and depth tolerances, and votes with depth-dependent weights into an occupancy grid. The reported numbers on NVOS (0.909 mean IoU) and Mip-NeRF360 (0.913 mIoU, tied with SANeRF-HQ) are the evidence offered for this claim. The paper also argues that this design removes the dependence of runtime on training-set size and enabl

What carries the argument

The load-bearing mechanism is the depth-weighted voxel aggregation: (i) inverse-normalized NeRF depth is multiplied into each SAM confidence map to suppress background and sharpen boundaries; (ii) a CUDA kernel projects each voxel to all views, checks density, ray-voxel distance tolerance τ_spatial, and depth tolerance τ_depth, and accumulates depth-weighted votes; (iii) a thin-structure path projects voxel corners to a 2D footprint and gives partial credit when a sufficient fraction of pixels support the voxel. This replaces the per-scene optimization loop with a single, fixed fusion step.

Load-bearing premise

The parity claim on Mip-NeRF360 rests on pseudo-ground-truth masks released by SANeRF-HQ being a fair yardstick, and the whole pipeline assumes the pre-trained NeRF's density and depth are reliable; the paper itself concedes both are imperfect in places.

What would settle it

Re-evaluate DivAS, SA3D, and SANeRF-HQ on Mip-NeRF360 using manually labeled ground-truth masks instead of the pseudo-GT from SANeRF-HQ; if the reported mIoU tie (0.913 vs 0.913) does not survive, the central parity claim is not independently established.

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

If this is right

  • Per-scene optimization can be replaced by a fixed fusion step: DivAS completes a typical scene in 4-6 minutes versus about 10 minutes for SANeRF-HQ, with compute largely independent of training-set size.
  • Zero-shot 2D segmentation transfers to 3D without retraining: SAM's masks, refined by NeRF depth, yield the highest mean IoU on NVOS (0.909) among compared methods.
  • Fine structures like ribs and wires can be segmented via a dedicated thin-structure path that gives partial credit for voxels with small 2D footprints, recovering details optimization-based methods miss.
  • The pipeline supports progressive refinement in a GUI: asynchronous SAM inference overlaps with user interaction, and the fusion kernel runs in under 200 ms, enabling real-time feedback.
  • Removing the thin-structure path drops IoU sharply on thin-geometry scenes (Trex 0.869 to 0.673), showing that the footprint check is load-bearing for fine details.

Where Pith is reading between the lines

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

  • The paper's broader abstract claim that the recipe transfers to Gaussian Splatting is not supported by the body, which evaluates only the NeRF instantiation; a fair test would be to run the same skeleton on a GS backbone and compare against optimization-based GS baselines.
  • The depth-weighting ablation shows almost no quantitative change on four scenes (e.g., 0.869 vs 0.867 on Trex) while the paper credits it with qualitative gains; this suggests the measured IoU advantage may come more from centroid-view zoom and thin-structure handling than from the depth prior itself.
  • The Mip-NeRF360 tie rests on pseudo-ground-truth masks from SANeRF-HQ; if those masks encode SANeRF-HQ's own biases, the parity result is weaker than it appears. A manual ground-truth evaluation would settle this.
  • The view-selection ablation (Fibonacci vs manual, 0.913 vs 0.920) suggests human guidance still helps in non-object-centric scenes; a usability study could quantify how much interactive effort is needed to reach the reported quality.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 5 minor

Summary. The manuscript presents DivAS, an optimization-free interactive 3D segmentation pipeline for NeRF scenes. The user prompts SAM on a small set of anchor and centroid views; SAM masks are computed asynchronously, refined by inverse-normalized NeRF depth, and fused into a 3D occupancy grid by a custom CUDA voxel kernel enforcing density, distance, and depth consistency. The paper reports mean IoU 0.909 on NVOS/LLFF (Table 4), mIoU 0.913 on Mip-NeRF360 (Table 5), and a 2-2.5x end-to-end speedup over SANeRF-HQ. The full-text submission is NeRF-specific; the arXiv metadata abstract additionally claims a Gaussian-Splatting instantiation, a consumer-hardware memory advantage, and a fusion-kernel cost below 70 ms, none of which appears in the body.

Significance. The core design is attractive and the fusion math in Eqs. (6)-(7) is a normalized weighted average with no fitted target, so there is no definitional circularity in the pipeline. If the headline results were supported, the paper would be a useful demonstration that a training-free fusion recipe can compete with per-scene optimization while preserving SAM's zero-shot behavior. However, the parity claim currently rests on one benchmark with genuine ground truth (six forward-facing scenes, no error bars) and one benchmark whose ground-truth masks were released by the main optimization baseline. The hyperparameter choice is tuned on the same scenes later scored, and the full depth-gradient formula and Table 1 hyperparameter values are deferred to an absent supplementary. These are fixable evaluation and reproducibility gaps, but they cap the strength of the claims as written.

major comments (5)
  1. [§4.1, Table 5] For Mip-NeRF360, the paper states that 'no true mask exist, so we adopt the pseudo-ground truth masks released by [17]' where [17] is SANeRF-HQ, the main optimization baseline. Table 5 then reports a tie (0.913 vs 0.913) between DivAS and SANeRF-HQ on these pseudo-masks. If the pseudo-GT encodes SANeRF-HQ's segmentation biases, this comparison is not independent evidence; §4.3 itself concedes that the pseudo-GT masks 'occasionally penalize our IoU'. Please evaluate on independently created masks or on a benchmark with real annotations, and separate the reported claim from qualitative examples.
  2. [§4.2, Fig. 2 vs Tables 4-5] The cumulative-weight threshold tau_cw is selected after an ablation on Trex, Fern, Bonsai, and Garden, and those same scenes later appear in the main results of Tables 4-5. Choosing tau_cw = 0.75 with this protocol is a form of test-set tuning. The authors should validate the threshold on held-out scenes or report the full sensitivity curve and justify the default without consulting the scoring set.
  3. [§3.5.3, Table 1] The depth-gradient factor g_i^p is defined only by saying 'Its full expression is included in the supplementary', no supplementary is provided. Table 1 lists hyperparameters gamma, beta, b_max, lambda_range, rho_thresh, rho_thin_thresh, and rho_percent_cover_thresh but gives no values. Consequently Eqs. (4)-(5) and the thin-structure coverage test cannot be evaluated or reimplemented from the manuscript. Please include the exact formula and a complete default-value table, or release code from which these values are read.
  4. [Abstract vs §4.5] The metadata abstract claims a representation-agnostic recipe instantiated on both Gaussian Splatting and NeRF, with the GS version 'the only one to reach this quality within the consumer-hardware memory envelope' and a 'fusion-kernel cost below 70 ms'. The body contains no GS experiments and §4.5 reports a 150-200 ms CUDA fusion kernel for the NeRF instantiation. The abstract and title should be aligned with the actual submission, or the GS experiments and memory measurements must be included.
  5. [Table 4] The NVOS comparison uses six forward-facing scenes with a single held-out view per scene and reports no error bars, repeated runs, or significance test. The mean IoU difference 0.909 vs 0.902 is small relative to per-scene variation; for example, DivAS is worse than SA3D on Fern (0.803 vs 0.822) and better on Orchids (0.895 vs 0.880). The claim of 'highest mean IoU' is therefore statistically fragile. Report variance over multiple click/prompt runs and a paired test.
minor comments (5)
  1. [§3] The text says 'As shown in Sec. 2.3, the pipeline consists of two stages'; this should refer to Figure 1 or Section 3, not Section 2.3.
  2. [Table 2] The column headers 'Thin: w', 'Thin: w/o', 'Depth: w', 'Depth: w/o' are ambiguous. Please spell out what each column includes and excludes, e.g., 'thin-structure path on/off' and 'depth weighting on/off'.
  3. [§3.4, Eq. (2)] The inverse normalized depth (Z - Zmin)/(Zmax - Zmin) can divide by zero for constant-depth regions; consider an epsilon or explicitly state how degenerate depth intervals are handled.
  4. [Supplementary] The manuscript repeatedly refers to a supplementary for procedural details, qualitative results, extended ablations, timing details, and the depth-gradient expression. No supplementary file is present in the arXiv submission; please include it in the next revision.
  5. [Abstract] Typographical issues: 'We introduceDivAS' is missing a space, and 'In zero-shot 3D segmentation' has inconsistent capitalization. The degree symbols in 'Mip-NeRF360 ◦' are also malformed in several places.

Circularity Check

1 steps flagged

τ_cw hyperparameter selected on the test scenes that are later scored; reported parity partly reflects this fit.

specific steps
  1. fitted input called prediction [§4.2 (Figure 2, τ_cw selection) and §4.3 Tables 4–5]
    "We evaluate τcw on scenes with narrow (Trex) and wide (Fern) depth ranges from LLFF, and on Bonsai and Garden from Mip-NeRF360◦. As shown in Fig. 2, IoU remains stable for τcw ∈ [0.3,0.75], while large values (≥0.9) sharply degrade performance. ... We therefore use τcw = 0.75 as a stable choice across datasets."

    The cumulative-weight cutoff τ_cw is a geometric gate in the fusion kernel (Eqs. 4–5) chosen after measuring IoU on Trex, Fern, Bonsai, and Garden. Those exact scenes are then scored in Table 4 (Trex, Fern) and Table 5 (Bonsai, Garden). The reported 'highest mean IoU' (0.909) and 'matches SANeRF-HQ in quantitative accuracy' (0.913 mIoU) are therefore not independent predictions: the threshold was selected to maximize/clean the IoU on the same evaluation scenes, so the final numbers partly restate the selection criterion rather than demonstrate zero-shot, parameter-free performance.

full rationale

The core DivAS fusion is not definitionally circular: Eq. (2) is a fixed elementwise transform of SAM confidence and normalized depth, and Eq. (7) is a normalized weighted average with no fitted target. The main circularity concern is in §4.2: the cumulative-weight cutoff τ_cw is selected after measuring IoU on Trex, Fern, Bonsai, and Garden, and those same four scenes appear in the headline NVOS (Table 4) and Mip-360 (Table 5) results. The reported 0.909 mean IoU and 0.913 mIoU therefore incorporate a test-set-informed choice of a geometric gate, so the 'optimization-free parity' claim is partly a restatement of that choice rather than a fully independent prediction. The Mip-360 benchmark also uses pseudo-ground-truth masks released by SANeRF-HQ [17], which weakens the independence of the comparison against SANeRF-HQ, though the paper's own admission that these masks 'occasionally penalize our IoU' indicates any bias runs against DivAS rather than for it. I do not treat the missing depth-gradient expression (deferred to an absent supplementary) or the valueless hyperparameters in Table 1 as circularity; they are completeness/reproducibility defects. There is no load-bearing self-citation chain: references [5], [17], etc. are external prior works. Score 4 reflects one moderate evaluation-set fit while the core derivation remains independent.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

All load-bearing constants are hand-chosen or fitted: τ_cw=0.75 via ablation on the evaluation scenes; zoom factors 0.67/0.47; ranking weights 0.4/0.3/0.3; ρ_thin=0.6; and the Table 1 tolerances whose values are not even disclosed. No invented entities. The heavy assumptions are domain-level: SAM reliability, Instant-NGP depth/density quality (explicitly acknowledged in Limitations), and the asserted behavior of the ad hoc depth-gradient factor, whose formula is deferred to a missing supplementary.

free parameters (6)
  • τ_cw (cumulative-weight cutoff) = 0.75
    Chosen via IoU sweep on Trex, Fern, Bonsai, Garden (§4.2, Fig. 2) — the same scenes that appear in the final evaluation Tables 4 and 5. Tuning on the evaluation set.
  • Kernel tolerances γ, β, b_max, λ_range, ρ_thresh, ρ_thin_thresh, ρ_percent_cover_thresh = unreported
    Table 1 lists purposes and says 'fixed per dataset' but gives no numeric values; the depth tolerance τ_depth = (γ + β·N_samples ≤ b_max)·Δx and spatial tolerance Eq. (4) depend on them.
  • α₁ (depth-confidence falloff) = unreported
    Eq. (6), w_depth = exp(−α₁·r²); paper asserts normalization makes it insensitive, but no value is given.
  • Zoom factors = 0.67 (LLFF), 0.47 (Mip-NeRF360)
    §3.2, hand-chosen per dataset for centroid-view generation; 'interactively adjustable'.
  • View ranking weights = 0.4 diversity, 0.3 coverage, 0.3 pitch
    §3.1, I(i) = 0.4D_i + 0.3C_i + 0.3P_i, hand-chosen.
  • ρ_thin (thin-structure acceptance threshold) = 0.6
    §3.6 compact fusion; a single strong mask activation above 0.6 is treated as evidence of object presence.
axioms (5)
  • domain assumption SAM's point-prompted zero-shot masks are a reliable 2D segmentation prior for arbitrary objects in rendered NeRF views
    Invoked throughout §3.3-3.5; the whole pipeline inherits SAM's failure modes, acknowledged in Limitations: 'recovering from all failure cases remains challenging.'
  • domain assumption NeRF max-weight-sample depth is a sharp surface estimate usable as a geometric gate
    §3.2 uses it for centroid-view pose; §3.4 uses inverse-normalized depth as the refinement weight in Eq. (2); Limitations concede degradation 'in scenes with poor geometry.'
  • domain assumption The density grid and voxel-space Euclidean projection with contraction yield view-consistent occupancy
    §3.5.2-3.5.3: voxel-to-view projection plus density threshold ρ_thresh define foreground validity; accuracy hinges on the pre-trained Instant-NGP NeRF's density quality.
  • ad hoc to paper Depth-gradient factor g^p behaves as described ('large on flat regions, small near boundaries')
    §3.5.3, Eq. (4): the tolerance τ_spatial depends on g^p, but its full expression is deferred to a supplementary that is not present, so its actual behavior is asserted rather than checkable.
  • standard math Fibonacci lattice gives low-discrepancy, uniform view coverage for small N
    §3.1, cited to [12]; standard geometry result, low risk.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DivAS: Interactive 3D Segmentation by Depth-Weighted Voxel Aggregation." pith.science (2026). https://pith.science/paper/MBESLMJA

@misc{pith2026260104860,
  author       = {Pith},
  title        = {Pith review of: DivAS: Interactive 3D Segmentation by Depth-Weighted Voxel Aggregation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBESLMJA}},
  note         = {Machine review of arXiv:2601.04860}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Interactive 3D segmentation of a reconstructed scene should not require a representation-specific optimization loop. We observe that the recipe for lifting 2D foundation-model masks into 3D, namely prompting a few views, refining the resulting masks with rendered depth, and fusing the multi-view evidence into a voxel grid, is shared across scene representations. What remains representation-specific is only the depth signal returned by the renderer and the occupancy prior that gates fusion. We present **DivAS** (Depth-interactive Voxel Aggregation Segmentation), an optimization-free, training-free framework that realizes this recipe as a single interaction-and-fusion skeleton with lightweight, representation-specific adapters, instantiated on both Gaussian Splatting (GS) and NeRF backbones. On standard forward-facing and unbounded benchmarks, the GS instantiation attains segmentation quality competitive with state-of-the-art optimization-based methods, and the best on LLFF, while being the only one to reach this quality within the consumer-hardware memory envelope at standard resolution. Both instantiations run end-to-end around $2$x faster than feature-field baselines, with a per-update fusion-kernel cost below $70$ ms. Because segmentation evidence is gathered from a small, bounded set of anchor views, user effort and computation remain independent of the training-set size. The same skeleton applied to a NeRF backbone matches or exceeds the performance of optimization-based NeRF baselines, confirming that the recipe transfers across fundamentally different 3D representations.

Figures

Figures reproduced from arXiv: 2601.04860 by Ayush Pande, Mayank Vatsa.

Figure 1
Figure 1. Figure 1: Overview of DivAS. (a) Stage 1: Anchor and centroid view generation. The user provides point prompts on global-scale anchor views to select distinct object regions. Each selected point generates a zoomed-in novel centroid view C by moving the camera toward the point and setting it as the look-at target. (b) Stage 2: Interactive centroid view processing and 3D fusion. For each centroid view cj in an anchor … view at source ↗
Figure 2
Figure 2. Figure 2: IoU vs cumulative-weight(τcw) target Cumulative weight(τcw): We evaluate τcw on scenes with narrow (Trex) and wide (Fern) depth ranges from LLFF, and on Bonsai and Garden from Mip-NeRF 360◦ . As shown in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Results of different methods on LLFF [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results of different methods on MIP-360◦ In [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

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

32 extracted references · 3 linked inside Pith

  1. [1]

    Segnet: A deep convolutional encoder-decoder architecture for image segmentation.IEEE transactions on pattern anal- ysis and machine intelligence, 39(12):2481–2495, 2017

    Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation.IEEE transactions on pattern anal- ysis and machine intelligence, 39(12):2481–2495, 2017. 2

    2017

  2. [2]

    Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields

    Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neu- ral radiance fields. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 5855–5864,

  3. [3]

    Mip-nerf 360: Unbounded anti-aliased neural radiance fields

    Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 1, 6

    2022

  4. [4]

    Emerg- ing properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv ´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 9650–9660, 2021. 2

    2021

  5. [5]

    Seg- ment anything in 3d with nerfs.Advances in Neural Infor- mation Processing Systems, 36:25971–25990, 2023

    Jiazhong Cen, Zanwei Zhou, Jiemin Fang, Wei Shen, Lingxi Xie, Dongsheng Jiang, Xiaopeng Zhang, Qi Tian, et al. Seg- ment anything in 3d with nerfs.Advances in Neural Infor- mation Processing Systems, 36:25971–25990, 2023. 1, 2, 7

    2023

  6. [6]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InEuropean con- ference on computer vision, pages 333–350. Springer, 2022. 2

    2022

  7. [7]

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs.IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017. 2

    2017

  8. [8]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 2

    Pith/arXiv arXiv 2010

  9. [9]

    Plenoxels: Radiance fields without neural networks

    Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2

    2022

  10. [10]

    Fastnerf: High-fidelity neu- ral rendering at 200fps

    Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neu- ral rendering at 200fps. InProceedings of the IEEE/CVF international conference on computer vision, pages 14346– 14355, 2021. 2

    2021

  11. [11]

    Interactive segmentation of radiance fields

    Rahul Goel, Dhawal Sirikonda, Saurabh Saini, and PJ Narayanan. Interactive segmentation of radiance fields. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4201–4211, 2023. 1, 2

    2023

  12. [12]

    Measurement of areas on a sphere using fibonacci and latitude–longitude lattices.Mathematical geo- sciences, 42(1):49–64, 2010

    ´Alvaro Gonz´alez. Measurement of areas on a sphere using fibonacci and latitude–longitude lattices.Mathematical geo- sciences, 42(1):49–64, 2010. 3

    2010

  13. [13]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. InProceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017. 2

    2017

  14. [14]

    Lerf: Language embedded radiance fields

    Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Lerf: Language embedded radiance fields. InProceedings of the IEEE/CVF interna- tional conference on computer vision, pages 19729–19739,

  15. [15]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 4015–4026, 2023. 1, 2, 3, 4

    2023

  16. [16]

    Decomposing nerf for editing via feature field distil- lation.Advances in neural information processing systems, 35:23311–23330, 2022

    Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitz- mann. Decomposing nerf for editing via feature field distil- lation.Advances in neural information processing systems, 35:23311–23330, 2022. 2

    2022

  17. [17]

    Sanerf-hq: Segment anything for nerf in high quality

    Yichen Liu, Benran Hu, Chi-Keung Tang, and Yu-Wing Tai. Sanerf-hq: Segment anything for nerf in high quality. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3216–3226, 2024. 2, 4, 6, 7, 8

    2024

  18. [18]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012–10022, 2021. 2

    2021

  19. [19]

    Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines.ACM Transac- tions on Graphics (ToG), 38(4):1–14, 2019

    Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view syn- thesis with prescriptive sampling guidelines.ACM Transac- tions on Graphics (ToG), 38(4):1–14, 2019. 3, 6

    2019

  20. [20]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. InEuropean Conference on Computer Vision, pages 405–421. Springer, 2020. 1, 2

    2020

  21. [21]

    Instant neural graphics primitives with a multires- olution hash encoding

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding. 2022. 2, 5

    2022

  22. [22]

    Langsplat: 3d language gaussian splatting

    Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3d language gaussian splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051–20060, 2024. 2

    2024

  23. [23]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 2

    2021

  24. [24]

    Neural volumetric object selection

    Zhongzheng Ren, Aseem Agarwala, Bryan Russell, Alexan- der G Schwing, and Oliver Wang. Neural volumetric object selection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6133–6142,

  25. [25]

    Segmenter: Transformer for semantic segmenta- tion

    Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmenta- tion. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 7262–7272, 2021. 2 9

    2021

  26. [26]

    Neural feature fusion fields: 3d distillation of self-supervised 2d image representations

    Vadim Tschernezki, Iro Laina, Diane Larlus, and Andrea Vedaldi. Neural feature fusion fields: 3d distillation of self-supervised 2d image representations. In2022 Inter- national Conference on 3D Vision (3DV), pages 443–453. IEEE, 2022. 2

    2022

  27. [27]

    Nto3d: Neural target object 3d reconstruction with segment anything

    Xiaobao Wei, Renrui Zhang, Jiarui Wu, Jiaming Liu, Ming Lu, Yandong Guo, and Shanghang Zhang. Nto3d: Neural target object 3d reconstruction with segment anything. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 20352–20362, 2024. 2

    2024

  28. [28]

    Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers.Advances in neural information processing systems, 34: 12077–12090, 2021. 2

    2021

  29. [29]

    Efficientsam: Leveraged masked image pretraining for efficient segment anything

    Yunyang Xiong, Bala Varadarajan, Lemeng Wu, Xiaoyu Xi- ang, Fanyi Xiao, Chenchen Zhu, Xiaoliang Dai, Dilin Wang, Fei Sun, Forrest Iandola, et al. Efficientsam: Leveraged masked image pretraining for efficient segment anything. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 16111–16121, 2024. 2

    2024

  30. [30]

    Faster segment anything: Towards lightweight sam for mo- bile applications.arXiv preprint arXiv:2306.14289, 2023

    Chaoning Zhang, Dongshen Han, Yu Qiao, Jung Uk Kim, Sung-Ho Bae, Seungkyu Lee, and Choong Seon Hong. Faster segment anything: Towards lightweight sam for mo- bile applications.arXiv preprint arXiv:2306.14289, 2023. 2

    Pith/arXiv arXiv 2023

  31. [31]

    In-place scene labelling and understanding with implicit scene representation

    Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and An- drew J Davison. In-place scene labelling and understanding with implicit scene representation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838–15847, 2021. 2

    2021

  32. [32]

    Medical sam 2: Segment medical images as video via segment anything model 2.arXiv preprint arXiv:2408.00874, 2024

    Jiayuan Zhu, Abdullah Hamdi, Yunli Qi, Yueming Jin, and Junde Wu. Medical sam 2: Segment medical images as video via segment anything model 2.arXiv preprint arXiv:2408.00874, 2024. 2 10

    Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.