Pith. sign in

REVIEW 4 major objections 5 minor 50 references

This paper proposes OSIM, an object-level metric for 3D scenes that matches human perceptual judgments more closely than existing whole-image and geometry metrics, and demonstrates this with user studies.

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

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

T0 review · deepseek-v4-flash

2026-08-04 19:35 UTC pith:7MHKGYBU

load-bearing objection A clean, well-defined object-centric 3D metric with useful diagnostics, but the claim that it beats all baselines on human correlation rests on a small, self-tuned user study. the 4 major comments →

arxiv 2509.09143 v1 pith:7MHKGYBU submitted 2025-09-11 cs.CV cs.AIcs.GR

Objectness Similarity: Capturing Object-Level Fidelity in 3D Scene Evaluation

Yuiko Uchida , Ren Togo , Keisuke Maeda , Takahiro Ogawa , Miki Haseyama This is my paper
classification cs.CV cs.AIcs.GR
keywords objectness similarityobject-centric evaluation3D scene evaluationnovel view synthesisperceptual quality metricsobject detection featuressaliency weightingmean opinion score correlation
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 paper tries to establish that evaluating a 3D scene by the quality of its individual objects, rather than by global pixel or feature statistics, yields scores that align more closely with human perception. It introduces Objectness SIMilarity (OSIM), which uses a pretrained object detector's intermediate features, compares the detected object patches between a rendered novel view and a reference view via cosine similarity, averages per object class, and weights the result by visual saliency. In user studies with 23 participants, OSIM achieves the highest rank correlation with mean opinion scores among all compared metrics, for both 3D reconstruction models and single-image-to-3D generation models. If the claim holds, OSIM provides a diagnostic tool that shows which objects in a scene are poorly reconstructed, and a standardized way to compare recent 3D methods.

Core claim

The central claim is that object-level feature similarity, computed from an object detection model, captures human perceptual fidelity in 3D scenes better than whole-image metrics. OSIM is computed as follows: for each novel-view render, detect objects in the reference image with YOLOX; crop the intermediate dark5 feature map at each bounding box; average the per-pixel cosine similarity between reference and test features inside the box; average these local scores over all instances of each object class; then take the saliency-weighted average over classes using GBVS saliency maps. The paper reports that OSIM reaches a rank correlation of 0.820 with MOS for reconstruction and 0.943 for gener

What carries the argument

The load-bearing mechanism is the intermediate feature map of a pretrained object detector, specifically the dark5 layer of YOLOX, combined with the detector's bounding-box proposals. These features are assumed to encode what is needed for category identification, so that feature similarity within a detected box is a proxy for whether the object remains recognizable as its class. Cosine similarity normalizes values to [0,1]; per-class averaging condenses instance-level scores; and GBVS saliency weighting makes the scene-level metric emphasize objects that naturally attract human attention.

Load-bearing premise

The metric assumes that the dark5 features of the pretrained YOLOX detector capture the same object-category cues that humans use when judging 3D scene quality, so that feature similarity inside detected boxes is a valid proxy for perceived objectness.

What would settle it

Take a dataset of scenes with objects outside YOLOX's training classes, or shuffle the dark5 features with random vectors of the same dimension; if OSIM's rank correlation with MOS drops to or below the level of whole-image metrics, then the specific object-detection features, rather than the object-centric localization, are carrying the result.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If OSIM is right, 3D reconstruction and generation should be evaluated on object-level fidelity in addition to pixel-level and geometry-level metrics, because the object-centric axis is the one that tracks human ratings.
  • OSIM can be used as a diagnostic: it pinpoints which object classes in a scene are poorly reconstructed, as demonstrated by the stop-sign and potted-plant examples.
  • The re-evaluation under a unified protocol suggests that recent 3D Gaussian splatting methods improved over NeRF-based methods, but that within the 2023–2024 splatting family there are no dramatic gains, so claims of progress should be read with standardized baselines.
  • Metrics computed on detected object patches, not just OSIM, correlate better with human perception than whole-image computations, implying that object-centric evaluation is broadly useful, not a quirk of the proposed formula.
  • OSIM's design extends naturally to 2D image quality assessment and to 4D dynamic scenes or video, which the paper names as future work.

Where Pith is reading between the lines

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

  • The reliance on a fixed set of detector classes means OSIM silently ignores objects outside YOLOX's category vocabulary; a natural test would be to swap in an open-vocabulary detector and see whether the correlation with MOS improves or degrades.
  • The saliency weighting is a free parameter choice; replacing GBVS with task-specific importance (e.g., a robot's manipulation targets) could make the metric application-dependent, which would be desirable for downstream uses.
  • The fact that patch-based versions of existing metrics also help suggests that OSIM's advantage might come partly from focusing computation on objects, partly from detector features; ablating those two contributions separately would isolate the source.
  • If the dark5 features are, in effect, a compressed representation of category identity, then OSIM might also serve as a training objective or a reward signal for iterative object-level refinement of reconstructed scenes.

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

4 major / 5 minor

Summary. The paper proposes OSIM, an object-centric evaluation metric for 3D scenes. OSIM uses a pre-trained YOLOX detector to locate objects in reference and novel-view images, extracts intermediate (dark5) features, computes per-object cosine similarities between corresponding bounding-box regions, and finally aggregates class-level scores weighted by saliency. The authors claim that OSIM better aligns with human perception than existing metrics (PSNR, SSIM, LPIPS, CLIP-sim., FID, CD, F-score) based on user studies, and they use OSIM to re-evaluate recent reconstruction and generation models under a standardized protocol.

Significance. If the central claim is established, OSIM would be a useful complement to existing full-image and geometry metrics, especially for diagnosing which objects in a scene are poorly reconstructed. The metric definition is simple, normalized, and makes the object-centric perspective explicit. The authors also contribute a comparative evaluation of many recent methods using consistent conditions, and they provide code. However, the empirical validation of the headline claim—that OSIM correlates better with human perception—is statistically underpowered and partly based on selective analysis, so the current strength of the evidence is limited.

major comments (4)
  1. [Sec. 4.3, Tables 2 and 3] The central claim that OSIM 'achieves the highest rank correlation among all metrics' is not supported by any significance testing, confidence intervals, or inter-rater reliability. For generation models, ρ is computed across only six models; the difference between OSIM (0.943) and CD (0.771) or CLIP-sim./FID (0.714) is plausibly within sampling noise at N=6. For reconstruction, the user study collects only 230 ratings over 84 scene-model pairs (~2.7 ratings per item), and no per-participant variance is reported. Please provide bootstrap CIs, permutation tests, or per-participant correlations, and account for the non-independence of ratings from the same participant and same scene.
  2. [Sec. 4.1, Table 5] The detection confidence threshold (0.35) is said to be chosen 'based on empirical results,' and the dark5 feature layer is selected after observing that it gives the highest rank correlation in Table 5. Both choices are made on the same evaluation data used to report OSIM's superiority. This constitutes selection over hyperparameters and can inflate the reported correlation. Please validate OSIM with held-out scenes/models or use nested cross-validation, and report the sensitivity of the main result to the threshold and layer choice.
  3. [Sec. 4.2.1 and Sec. 4.3] COLMAP is excluded from the correlation analyses and from Fig. 7 because its scores are 'extremely low' and treated as an outlier. This exclusion is post-hoc and no objective outlier criterion is given. Since Spearman rank correlations over a small number of models can be strongly affected by a single extreme point, report the correlations both with and without COLMAP, and either pre-specify an outlier rule or show that the conclusion is robust to its inclusion.
  4. [Sec. 4.3] The user study is conducted with 23 participants from a single research lab, and no inter-rater reliability or consistency measure is reported. The rating instructions explicitly ask participants to consider 'objectness' and semantic fidelity, which may steer them toward the same cues that OSIM uses; the comparison with metrics that were not designed with this instruction is therefore not fully controlled. Please report inter-rater agreement and, if possible, validate on an external or pre-existing subjective dataset, or at least discuss this limitation and its potential effect on the ranking.
minor comments (5)
  1. [Sec. 4.3] Typo: 'creiterion' should be 'criterion'.
  2. [Related Work and Table 3] Formatting: 'V olume' appears with an odd spacing; please use 'Volume' consistently.
  3. [Fig. 1] The top panel shows numerical values (e.g., 26.363, 0.904) that are not explained in the caption or text. Please clarify which metric each number corresponds to.
  4. [Table 2] The table reports 'Average Spearman’s rank correlation' but gives a single value per metric. Clarify what the average is over (e.g., scenes? participants?) and provide variability across bootstrap samples or scenes.
  5. [Sec. 4.4] The normalization of FID in Fig. 7 ('both division and subtraction from 1') is unclear. Please specify the exact transformation and why it was applied.

Circularity Check

1 steps flagged

OSIM's reported perceptual correlation is partially a selected maximum: the feature layer (and threshold) were chosen using the same MOS data on which the top-rank claim is made.

specific steps
  1. fitted input called prediction [Sec 4.1 (Experimental Settings) and Sec 5 (Ablation Study), Tables 2, 3 and 5]
    "We set the detection confidence threshold to 0.35 based on empirical results and adopted the original settings for the rest. The intermediate features were derived from the dark5 layer of the backbone since it represents global image context and contains information necessary for category identification. ... The results show that the highest rank correlation is achieved with dark5, while p5 also tends to yield relatively better performance."

    The paper's central validation is the claim that OSIM 'achieves the highest rank correlation' with MOS (Sec 4.3, Tables 2 and 3). But the OSIM configuration used for those correlations was selected from the same MOS data: Sec 4.1 says the detection threshold was set 'based on empirical results,' and Sec 5's ablation sweeps six intermediate layers (dark3, dark4, dark5, p3, p4, p5) and reports rank correlations on the same Reconstruction and Generation datasets, with dark5 giving the highest values (0.820 and 0.943). Reporting the maximum over these model-selection choices as evidence of OSIM's superiority is a fitted-input-called-prediction circularity: the advantage may reflect tuning the layer/threshold to the target MOS rather than an independent predictive property. The metric itself is

full rationale

The core OSIM computation is self-contained: it uses the pre-trained YOLOX-x detector, GBVS saliency, and the equations in Secs 3.3-3.4, none of which are fit to the user-study MOS labels. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled via citation. However, the paper's headline empirical claim—that OSIM aligns more closely with human perception—rests on Spearman correlations obtained after choosing the detection threshold and feature layer on the same evaluation data. The Table 5 ablation explicitly selects dark5 because it maximizes rank correlation on these datasets, and the reported Tables 2-3 values are therefore the outcome of a selection procedure, not an out-of-sample prediction. This is a genuine fitted-input-called-prediction circularity, though it does not make the metric definition itself circular. The additional statistical weaknesses (N=6 for generation, no confidence intervals or significance tests) are real but are correctness/statistical-power concerns rather than circularity, and are noted here only for completeness. Overall score 6 reflects one central partially circular step while acknowledging the metric has substantial independent content.

Axiom & Free-Parameter Ledger

1 free parameters · 3 axioms · 0 invented entities

The metric leans on three external assumptions: object-based attention, saliency transfer, and detection-feature validity. No new physical entities are introduced. The main free parameter is the detection confidence threshold.

free parameters (1)
  • Detection confidence threshold = 0.35
    Set in Sec 4.1 'based on empirical results'; affects which detected objects are included in OSIM and, in turn, the final score.
axioms (3)
  • domain assumption Human scene perception is object-based, prioritizing coherent bounded objects (Scholl 2001)
    Motivates the entire objectness criterion in the Introduction; if object attention is not a dominant factor, the metric's premise weakens.
  • domain assumption Saliency maps computed on reference images approximate the attention weights for objects in novel-view images
    Used in Sec 3.4 to weight object class scores; reference-image saliency is assumed to transfer across views.
  • domain assumption Pre-trained YOLOX dark5 features encode object-category information that degrades monotonically with rendering quality
    Sec 3.2 states these features contain essential information for object identification; Sec 4.1 selects dark5 as the best layer. This underpins the metric's validity.

pith-pipeline@v1.3.0-alltime-deepseek · 14284 in / 8954 out tokens · 98894 ms · 2026-08-04T19:35:04.846801+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Objectness Similarity: Capturing Object-Level Fidelity in 3D Scene Evaluation." pith.science (2026). https://pith.science/paper/7MHKGYBU

@misc{pith2026250909143,
  author       = {Pith},
  title        = {Pith review of: Objectness Similarity: Capturing Object-Level Fidelity in 3D Scene Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7MHKGYBU}},
  note         = {Machine review of arXiv:2509.09143}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper presents Objectness SIMilarity (OSIM), a novel evaluation metric for 3D scenes that explicitly focuses on "objects," which are fundamental units of human visual perception. Existing metrics assess overall image quality, leading to discrepancies with human perception. Inspired by neuropsychological insights, we hypothesize that human recognition of 3D scenes fundamentally involves attention to individual objects. OSIM enables object-centric evaluations by leveraging an object detection model and its feature representations to quantify the "objectness" of each object in the scene. Our user study demonstrates that OSIM aligns more closely with human perception compared to existing metrics. We also analyze the characteristics of OSIM using various approaches. Moreover, we re-evaluate recent 3D reconstruction and generation models under a standardized experimental setup to clarify advancements in this field. The code is available at https://github.com/Objectness-Similarity/OSIM.

Figures

Figures reproduced from arXiv: 2509.09143 by Keisuke Maeda, Miki Haseyama, Ren Togo, Takahiro Ogawa, Yuiko Uchida.

Figure 1
Figure 1. Figure 1: (Top): An example of evaluation results. Darker dots in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our proposed metric. First, we split the dataset and perform reconstruction or generation followed by rendering. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (left): Average Pearson correlation coefficients on the [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Object-level evaluation examples. (Top): Most objects [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of low-quality objects with semi [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Re-evaluation results on Mip-NeRF360 dataset and GSO dataset on our proposed metric. The models are arranged from left to [PITH_FULL_IMAGE:figures/full_fig_p008_7.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

50 extracted references · 6 linked inside Pith

  1. [1]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    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, 3, 4, 5, 6, 7, 8

    2022

  2. [2]

    Barron, Ben Mildenhall, Dor Verbin, Pratul P

    Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 1, 4, 7, 8

    2023

  3. [3]

    Distance transformations in digital im- ages.Computer Vision, Graphics, and Image Processing, 34 (3):344–371, 1986

    Gunilla Borgefors. Distance transformations in digital im- ages.Computer Vision, Graphics, and Image Processing, 34 (3):344–371, 1986. 1, 2, 5

    1986

  4. [4]

    Predicting human gaze using low-level saliency combined with face detection.Advances in Neural Information Processing Systems, 20, 2007

    Moran Cerf, Jonathan Harel, Wolfgang Einh ¨auser, and Christof Koch. Predicting human gaze using low-level saliency combined with face detection.Advances in Neural Information Processing Systems, 20, 2007. 5

    2007

  5. [5]

    Tensorf: Tensorial radiance fields

    Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. InProceedings of the European Conference on Computer Vision, pages 333– 350, 2022. 4

    2022

  6. [6]

    A survey on 3D gaussian splatting.arXiv:2401.03890, 2024

    Guikun Chen and Wenguan Wang. A survey on 3D gaussian splatting.arXiv:2401.03890, 2024. 1, 2

    Pith/arXiv arXiv 2024

  7. [7]

    Robbersmyr, and Kris- tian Muri Knausg ˚ard

    Anurag Dalal, Daniel Hagen, Kjell G. Robbersmyr, and Kris- tian Muri Knausg ˚ard. Gaussian splatting: 3D reconstruc- tion and novel view synthesis: A review.IEEE Access, 12: 96797–96820, 2024. 1, 2

    2024

  8. [8]

    McHugh, and Vincent Vanhoucke

    Laura Downs, Anthony Francis, Nate Koenig, Brandon Kin- man, Ryan Hickman, Krista Reymann, Thomas B. McHugh, and Vincent Vanhoucke. Google scanned objects: A high- quality dataset of 3D scanned household items. InProceed- ings of the International Conference on Robotics and Au- tomation, pages 2553–2560, 2022. 4, 6, 8

    2022

  9. [9]

    3D gaussian splatting as new era: A survey.IEEE Transactions on Visualization and Computer Graphics, pages 1–20, 2024

    Ben Fei, Jingyi Xu, Rui Zhang, Qingyuan Zhou, Weidong Yang, and Ying He. 3D gaussian splatting as new era: A survey.IEEE Transactions on Visualization and Computer Graphics, pages 1–20, 2024. 1, 2

    2024

  10. [10]

    K-planes: Explicit radiance fields in space, time, and appearance

    Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 12479–12488, 2023. 4

    2023

  11. [11]

    Nerf: Neural radiance field in 3D vision, a comprehensive review.arXiv:2210.00379, 2022

    Kyle Gao, Yina Gao, Hongjie He, Dening Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3D vision, a comprehensive review.arXiv:2210.00379, 2022. 1, 2

    Pith/arXiv arXiv 2022

  12. [12]

    Yolox: Exceeding yolo series in 2021

    Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv:2107.08430, 2021. 4

    Pith/arXiv arXiv 2021

  13. [13]

    Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in Neural Information Processing Systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- rium.Advances in Neural Information Processing Systems, 30, 2017. 2, 5, 7

    2017

  14. [14]

    2D gaussian splatting for geometrically ac- curate radiance fields

    Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically ac- curate radiance fields. InProceedings of the ACM SIG- GRAPH 2024 Conference Papers, pages 1–11, 2024. 4, 8

    2024

  15. [15]

    Brian K. S. Isaac-Medina, Chris G. Willcocks, and Toby P. Breckon. Exact-nerf: An exploration of a precise volumetric parameterization for neural radiance fields. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 66–75, 2023. 1

    2023

  16. [16]

    3D gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42 (4):1–14, 2023. 1, 3, 4, 8

    2023

  17. [17]

    3D gaussian splat- ting as markov chain monte carlo.Advances in Neural Infor- mation Processing Systems, 37:80965–80986, 2024

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Wei- wei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D gaussian splat- ting as markov chain monte carlo.Advances in Neural Infor- mation Processing Systems, 37:80965–80986, 2024. 4, 8

    2024

  18. [18]

    Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics, 36(4):1–13,

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction.ACM Transactions on Graphics, 36(4):1–13,

  19. [19]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural net- works.Advances in Neural Information Processing Systems, 25, 2012. 2, 5

    2012

  20. [20]

    Tetra-nerf: Represent- ing neural radiance fields using tetrahedra

    Jonas Kulhanek and Torsten Sattler. Tetra-nerf: Represent- ing neural radiance fields using tetrahedra. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 18458–18469, 2023. 4

    2023

  21. [21]

    Nerfbaselines: Con- sistent and reproducible evaluation of novel view synthesis methods.arXiv:2406.17345, 2024

    Jonas Kulhanek and Torsten Sattler. Nerfbaselines: Con- sistent and reproducible evaluation of novel view synthesis methods.arXiv:2406.17345, 2024. 2, 4, 8

    arXiv 2024

  22. [22]

    Advances in 3D generation: A survey

    Xiaoyu Li, Qi Zhang, Di Kang, Weihao Cheng, Yiming Gao, Jingbo Zhang, Zhihao Liang, Jing Liao, Yan-Pei Cao, and Ying Shan. Advances in 3D generation: A survey. arXiv:2401.17807, 2024. 2

    Pith/arXiv arXiv 2024

  23. [23]

    One-2-3-45++: Fast single im- age to 3D objects with consistent multi-view generation and 3D diffusion

    Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Ji- ayuan Gu, and Hao Su. One-2-3-45++: Fast single im- age to 3D objects with consistent multi-view generation and 3D diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072– 10083, 2024. 1, 4

    2024

  24. [24]

    Scaffold-GS: Structured 3D gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D gaussians for view-adaptive rendering. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 20654–20664, 2024. 4, 8

    2024

  25. [25]

    Srinivasan, Matthew Tancik, Jonathan T

    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.Communications of the ACM, 65(1):99–106, 2021. 1, 4 9

    2021

  26. [26]

    Instant neural graphics primitives with a multires- olution hash encoding.ACM Transactions on Graphics, 41 (4):1–15, 2022

    Thomas M ¨uller, Alex Evans, Christoph Schied, and Alexan- der Keller. Instant neural graphics primitives with a multires- olution hash encoding.ACM Transactions on Graphics, 41 (4):1–15, 2022. 1, 4, 8

    2022

  27. [27]

    Karl Pearson. Vii. note on regression and inheritance in the case of two parents.The Royal Society of London, 58(347- 352):240–242, 1895. 5

  28. [28]

    Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

    Alec Radford, Jong W. Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning trans- ferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748–8763, 2021. 2, 5, 7

    2021

  29. [29]

    Brian J. Scholl. Objects and attention: The state of the art. Cognition, 80(1):1–46, 2001. 1

    2001

  30. [30]

    Schonberger and Jan-Michael Frahm

    Johannes L. Schonberger and Jan-Michael Frahm. Structure- from-motion revisited. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 4104–4113, 2016. 4, 7, 8

    2016

  31. [31]

    Sch ¨onberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys

    Johannes L. Sch ¨onberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. InProceedings of the Euro- pean Conference on Computer Vision, pages 501–518, 2016. 4, 5, 7, 8

    2016

  32. [32]

    Sheikh, Muhammad F

    Hamid R. Sheikh, Muhammad F. Sabir, and Alan C. Bovik. A statistical evaluation of recent full reference image quality assessment algorithms.IEEE Transactions on Image Pro- cessing, 15(11):3440–3451, 2006. 2

    2006

  33. [33]

    Very deep convo- lutional networks for large-scale image recognition

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition. InPro- ceedings of the International Conference on Learning Rep- resentations, 2014. 2

    2014

  34. [34]

    The proof and measurement of associa- tion between two things.The American Journal of Psychol- ogy, 15(1):72–101, 1904

    Charles Spearman. The proof and measurement of associa- tion between two things.The American Journal of Psychol- ogy, 15(1):72–101, 1904. 6, 7, 8

    1904

  35. [35]

    Rethinking the inception archi- tecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception archi- tecture for computer vision. InProceedings of the IEEE Con- ference on Computer Vision and Pattern Recognition, pages 2818–2826, 2016. 2

    2016

  36. [36]

    Nerfstudio: A modular framework for neural radiance field development

    Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David Mcallister, Justin Kerr, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. InPro- ceedings of the ACM SIGGRAPH 2023 Conference Papers,

    2023

  37. [37]

    Dreamgaussian: Generative gaussian splatting for ef- ficient 3D content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for ef- ficient 3D content creation. InProceedings of the Interna- tional Conference on Learning Representations, 2023. 1, 4

    2023

  38. [38]

    LGM: Large multi-view gaussian model for high-resolution 3D content creation

    Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3D content creation. InProceed- ings of the European Conference on Computer Vision, pages 1–18, 2024. 1, 4

    2024

  39. [39]

    Richter, Ren ´e Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox

    Maxim Tatarchenko, Stephan R. Richter, Ren ´e Ranftl, Zhuwen Li, Vladlen Koltun, and Thomas Brox. What do single-view 3D reconstruction networks learn? InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3405–3414, 2019. 2, 5

    2019

  40. [40]

    Tri- posr: Fast 3D object reconstruction from a single image

    Dmitry Tochilkin, David Pankratz, Zexiang Liu, Zixuan Huang, Adam Letts, Yangguang Li, Ding Liang, Chris- tian Laforte, Varun Jampani, and Yan-Pei Cao. Tri- posr: Fast 3D object reconstruction from a single image. arXiv:2403.02151, 2024. 4

    Pith/arXiv arXiv 2024

  41. [41]

    Simoncelli, and Alan C

    Zhou Wang, Eero P. Simoncelli, and Alan C. Bovik. Mul- tiscale structural similarity for image quality assessment. In The Thrity-Seventh Asilomar Conference on Signals, Systems & Computers, 2003, pages 1398–1402. Ieee, 2003. 2

    2003

  42. [42]

    Bovik, Hamid R

    Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: From error visibility to structural similarity.IEEE Transactions on Image Pro- cessing, 13(4):600–612, 2004. 1, 2, 3, 5

    2004

  43. [43]

    CRM: Single image to 3D textured mesh with convolu- tional reconstruction model

    Zhengyi Wang, Yikai Wang, Yifei Chen, Chendong Xiang, Shuo Chen, Dajiang Yu, Chongxuan Li, Hang Su, and Jun Zhu. CRM: Single image to 3D textured mesh with convolu- tional reconstruction model. InProceedings of the European Conference on Computer Vision, pages 1–19, 2024. 4

    2024

  44. [44]

    Unique3D: High-quality and efficient 3D mesh generation from a single image.Advances in Neural Information Pro- cessing Systems, 37:125116–125141, 2024

    Kailu Wu, Fangfu Liu, Zhihan Cai, Runjie Yan, Hanyang Wang, Yating Hu, Yueqi Duan, and Kaisheng Ma. Unique3D: High-quality and efficient 3D mesh generation from a single image.Advances in Neural Information Pro- cessing Systems, 37:125116–125141, 2024. 4

    2024

  45. [45]

    gsplat: An open-source library for gaussian splatting.Journal of Machine Learning Research, 26(34):1–17, 2025

    Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting.Journal of Machine Learning Research, 26(34):1–17, 2025. 4, 8

    2025

  46. [46]

    Mip-splatting: Alias-free 3D gaussian splat- ting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D gaussian splat- ting. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 19447–19456,

  47. [47]

    Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes.ACM Transactions on Graphics, 43(6): 1–13, 2024

    Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes.ACM Transactions on Graphics, 43(6): 1–13, 2024. 4, 8

    2024

  48. [48]

    Efros, Eli Shecht- man, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shecht- man, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018. 1, 2, 5, 7

    2018

  49. [49]

    Evaluating human perception of novel view synthesis: Subjective quality as- sessment of gaussian splatting and nerf in dynamic scenes

    Yuhang Zhang, Joshua Maraval, Zhengyu Zhang, Nicolas Ramin, Shishun Tian, and Lu Zhang. Evaluating human perception of novel view synthesis: Subjective quality as- sessment of gaussian splatting and nerf in dynamic scenes. arXiv:2501.08072, 2025. 1, 6

    Pith/arXiv arXiv 2025

  50. [50]

    Zou, Zhipeng Yu, Yuan C

    Zi X. Zou, Zhipeng Yu, Yuan C. Guo, Yangguang Li, Ding Liang, Yan P. Cao, and Song H. Zhang. Triplane meets gaus- sian splatting: Fast and generalizable single-view 3D recon- struction with transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324–10335, 2024. 1, 4 10

    2024