Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

HairGS claims that strand-level hair geometry can be reconstructed from multi-view images by optimizing 3D Gaussians, merging segments with distance/angle heuristics, and refining under photometric supervision—no learned priors needed, in a

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 21:41 UTC pith:7D3FQLEF

load-bearing objection A practical, fast 3DGS hair pipeline whose speed and qualitative results are genuine, but the new 'strand consistency' metric doesn't measure what it claims and the baseline comparison is not fair. the 3 major comments →

arxiv 2509.07774 v1 pith:7D3FQLEF submitted 2025-09-09 cs.CV

HairGS: Hair Strand Reconstruction based on 3D Gaussian Splatting

Yimin Pan , Matthias Nie{\ss}ner , Tobias Kirschstein This is my paper
classification cs.CV
keywords hair reconstruction3D Gaussian splattingstrand mergingmulti-view reconstructionstrand consistencydifferentiable renderingdigital human modeling
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.

This paper claims that strand-level 3D hair geometry can be reconstructed from a small set of multi-view images by adapting 3D Gaussian splatting into a three-stage optimization: recover dense hair geometry as Gaussians, merge Gaussian segments into polylines using distance and direction thresholds, then refine joint positions under photometric and smoothness losses. The method deliberately avoids learned hair priors, relying instead on image-based supervision throughout, and the authors report that it reconstructs straight, curly, and long hairstyles in under an hour. To make topological quality visible, the paper introduces a strand-consistency score that measures what fraction of a ground-truth strand is covered by a single predicted strand. If the claim holds, strand-accurate hair capture becomes an optimization problem solvable on one consumer GPU rather than a data-hungry learning task.

Core claim

The central claim is that hair strand topology can be recovered without any pretrained hair model: a differentiable Gaussian rasterizer first fits a dense set of anisotropic Gaussians to the observed images; a greedy nearest-neighbour merge then connects close, similarly oriented endpoints into longer strands; a final photometric refinement plus angle-smoothness and adaptive point insertion turns those merged chains into accurate polylines. The paper reports that on the USC-HairSalon benchmark and on a curly Cem-Yuksel test model, this pipeline outperforms data-driven and classical baselines on precision, recall, F-score, and the new strand-consistency metric, while completing in roughly one

What carries the argument

The load-bearing mechanism is the merging scheme of Stage II: each first-stage Gaussian is seeded as a two-joint strand, and endpoints are matched greedily as nodes in a bipartite graph whose edge cost combines Euclidean distance and direction difference, implemented with a K-D tree and thresholds that relax from 2 mm and 20 degrees to 4 mm and 40 degrees during optimization. This geometric heuristic is the only mechanism that produces strand connectivity; the subsequent losses refine positions and topology but cannot create new links.

Load-bearing premise

The strand merging step assumes that two strand endpoints that lie within 2–4 mm and within 20–40 degrees of each other belong to the same physical strand, so a greedy nearest-neighbour match can recover true connectivity from geometry alone; the paper itself concedes this can fail in dense or curly hair, producing shorter reconstructed strands.

What would settle it

Take a dense curly hairstyle with known ground-truth strands, run only the merging stage on perfect Gaussian endpoints, and measure strand consistency as the distance and angle thresholds are relaxed past the paper's range; if SC saturates well below the pointwise F-score, the endpoint heuristic itself is the bottleneck. Conversely, if SC rises to match F-score, connectivity recovery is essentially solved and the short-strand limitation is merely a threshold-tuning issue.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Strand-level hair geometry can be produced from a small set of multi-view images in about one hour on a single consumer GPU, well below the 48–120 hours reported for learning-based baselines.
  • Because no learned hair prior is used, the method can in principle generalize to hairstyles outside synthetic training distributions, including curly and fine floating strands.
  • The strand-consistency metric provides a way to score topological accuracy rather than only pointwise geometric accuracy, filling a gap in existing precision/recall evaluations.
  • The output is a set of polylines with variable joint counts, so strand length and point density are determined by the optimization rather than by a fixed template.
  • Relaxing the merging thresholds during optimization approximately doubles the average reconstructed strand length, indicating that connectivity is a tunable trade-off against geometric precision.

Where Pith is reading between the lines

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

  • One testable consequence not explored in the paper: the strand-consistency metric could be used as an in-the-loop training signal or early-stopping criterion for the merging stage, rather than only as an offline benchmark.
  • A natural extension the paper hints at but does not test is transferring the same segmentation-plus-merging pipeline to other line-like structures such as cables, wires, or plant roots; the central assumption would still be that geometry alone can identify connectivity.
  • The reported strand-consistency ceiling (about 0.19 even under relaxed thresholds) suggests the geometric endpoint heuristic, not threshold choice, is the main bottleneck for dense curly hair; replacing greedy matching with a globally optimal assignment could be a direct next experiment.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes HairGS, a three-stage pipeline for strand-level hair reconstruction from multi-view images using 3D Gaussian Splatting (3DGS). Stage I optimizes a set of 3D Gaussians under photometric, orientation, and mask losses. Stage II converts each Gaussian into a two-joint strand and greedily merges endpoints using distance/angle thresholds, with iterative relaxation of these thresholds. Stage III refines joint positions under photometric supervision plus a smoothness loss and splits overly long segments. The paper reports quantitative results on USC-HairSalon and a Cem-Yuksel curly sample, comparing against Neural Haircut, LP-MVS, and Strand Integration, and introduces a new Strand Consistency (SC) metric as a proxy for topological accuracy. The abstract claims that the method handles a wide range of hairstyles and completes in about one hour.

Significance. If the claims held, HairGS would be a practical, prior-free alternative to learning-based hair reconstruction, with a new topology-aware evaluation metric. The paper's strengths are its clean ablation structure (Figure 5, Table 3), the explicit threshold sweep, the fully optimization-based design without learned hair priors, and the reported runtime under one hour. However, the central quantitative evidence for the topological/connectivity contribution rests on the proposed SC metric, which, as defined in Eq. (8), does not penalize over-merging and therefore does not measure what it claims. This undermines the paper's main comparative claim until the metric is reworked or corroborated by a precision-side connectivity measure.

major comments (3)
  1. [Section 4.2, Eq. (8)] The Strand Consistency (SC) metric is one-sided and does not measure topological accuracy. For each GT strand sG, SC takes the maximum over predicted strands sP of the fraction of sG's points matched into that single sP. A predicted strand that over-merges points from many different GT strands receives a high SC for each of those GT strands, because spurious connectivity is never penalized. Thus SC rewards long, coarsely merged connected components. The paper's interpretation in Section 4.3 — that higher SC means 'more reliable reconstruction of correct strand connectivity' — is not supported by Eq. (8). This is exactly the failure mode the paper itself concedes in Section 5 (merging can produce shorter strands) and that the relaxed thresholds in Section 3.3 (d_m from 2 to 4 mm, theta_m from 20 to 40 degrees) may create. The SC results in Tables 1-3 therefore do not substantiate the topo
  2. [Section 4.1, Baselines] The comparison against LP-MVS and Strand Integration is biased by the postprocessing. The paper states that for these methods, the directed point clouds are converted to strands using only the forward Euler method from [19] and that this 'lacks the mean-shift and growing steps which are essential for producing clean, long hair strands.' The reported P/R/F and SC for these baselines therefore reflect ablated, non-official variants rather than the methods as published. This makes the central claim of outperforming these methods on topology and overall quality unfair as stated. Please use the official strand-forming pipelines for those baselines, or clearly label the comparison as against incomplete variants and justify why this is a fair test.
  3. [Section 4.2, Metric validation] The proposed SC metric is not validated as a proxy for topological accuracy. There is no experiment showing that SC correlates with visually judged connectivity or with known ground-truth topology errors. Because SC ignores over-merging, the highest SC scores of the proposed method could simply reflect longer predicted strands rather than correct strand matching. The low absolute SC values (max 0.2196 in Table 3, 0.189 in Table 1, 0.166 in Table 2) also indicate that even by the paper's own metric, the reconstructed connectivity is far from correct. Please provide a validation experiment — for example, perturbing ground-truth strands by random splitting and merging and showing how SC changes — and discuss the absolute SC numbers more carefully.
minor comments (4)
  1. [Throughout] Typos: 'negleting' in Section 1, 'begining' in Section 2, and the header 'Spla tting' in the running title. Please proofread.
  2. [Eq. (5)] The Rodrigues formula notation is ambiguous: K is first defined as the cross product ⃗x × ⃗p_j, but the formula uses K as the skew-symmetric matrix. The denominator term '1+v·d' is not defined. Please clarify the notation.
  3. [Table 3, row 4] The F-score values in the row '6mm/60◦ →8mm/80◦' appear inconsistent with the reported precision and recall. For P=0.4355, R=0.1376, F1 should be approximately 0.209, not 0.271; for P=0.775, R=0.5182, F1 should be approximately 0.620, not 0.565. Please check and correct the table.
  4. [Section 4.1] The synthetic evaluation uses simple ambient/diffuse lighting and brown hair. The paper later attributes Neural Haircut's low scores to 'limited generalization to the synthetic evaluation data' (Section 4.3). Please state explicitly whether the synthetic rendering protocol matches Neural Haircut's training distribution, since the generalization claim depends on this.

Circularity Check

0 steps flagged

No significant circularity: the method is empirically benchmarked against external datasets and baselines, and the novel SC metric is an evaluation definition, not a fitted prediction or derivation.

full rationale

The paper's central claims are empirical comparisons, not analytic derivations. The multi-stage pipeline (Sec. 3) is evaluated on the external USC-HairSalon and Cem-Yuksel datasets against Neural Haircut, LP-MVS, and Strand Integration using precision, recall, and F-score, which are standard external metrics. The proposed strand-consistency metric (Eq. 8) is a new evaluation definition; even if it is a one-sided proxy that can reward over-merging, that is a question of metric validity, not circularity. No equation in the paper reduces to another by construction, no fitted parameter is renamed as a prediction, and no load-bearing claim rests on a self-citation. The only self-citations (NeRSemble [8], GaussianAvatars [21]) appear as a qualitative dataset and related-work context. The conclusion's admitted limitation that merging can yield shorter strands weakens the topology claim but does not make the derivation circular. The paper is self-contained against external benchmarks, so the appropriate score is 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The pipeline rests on five domain assumptions about preprocessing (COLMAP, FLAME, Gabor orientation, off-the-shelf masks), the view-independence of hair appearance (SH degree 0), and the validity of geometric endpoint merging as a topology oracle. Free parameters are the hand-chosen merging thresholds (2 mm and 20 degrees relaxed to 4 mm and 40 degrees), the unreported smoothness threshold and loss weights, and the fixed 30k plus 30k iteration schedule. The only invented construct is the Strand Consistency metric, a methodological artifact with no independent validation and a known blind spot for spurious connections.

free parameters (4)
  • Merging distance threshold d_m = 2 mm relaxed to 4 mm
    Hand-chosen (Sec. 3.2 and 3.3); directly controls how many merges occur and thus strand length and the SC metric. Ablation (Table 3) shows the 1 to 8 mm range is stable but extreme values degrade accuracy.
  • Merging angle threshold theta_m = 20 degrees relaxed to 40 degrees
    Hand-chosen endpoint direction agreement cutoff (Sec. 3.2 and 3.3); with d_m it defines the entire merging behavior and therefore the topology claim.
  • Smoothness angle threshold theta_s and loss weights lambda_theta, lambda_m, lambda_smooth, lambda_DSSIM = not reported
    Values are not given in the text (Eqs. 4, 6, 7); the orientation, mask, and refinement behavior depends on them.
  • Per-stage iteration counts = 30,000 + 30,000
    Fixed schedule (Sec. 4.1); densification and convergence behavior depends on it, and it drives the claimed one-hour runtime.
axioms (5)
  • domain assumption COLMAP recovers sufficiently accurate camera poses and a sparse point cloud for 3DGS initialization.
    Invoked in Sec. 3; the pipeline inherits 3DGS's known sensitivity to initialization.
  • domain assumption FLAME model fitting provides a good initialization point cloud.
    Sec. 3: 'we use the vertices obtained from fitting a FLAME model'; this import of a parametric head prior is an unstated dependency that weakens the 'no priors' framing.
  • domain assumption Gabor-filter orientation fields and off-the-shelf hair masks (MODNet, CDGNet) are reliable enough to supervise hair direction and isolate hair regions.
    Sec. 3; the orientation loss L_theta and mask loss L_m rest entirely on these preprocessing outputs, and errors propagate into the Gaussian geometry.
  • domain assumption Hair radiance is effectively view-independent (SH degree 0), so the photometric loss is a valid geometric supervisor.
    Sec. 3.1: SH degree 0 is chosen 'to focus on geometry rather than appearance'; real hair has specular, view-dependent appearance, so the assumption is convenient but not physically exact.
  • domain assumption Merging endpoints by proximity and direction recovers true strand connectivity.
    Sec. 3.2; the central topology mechanism. Admitted to fail in practice in the conclusion ('resulting in shorter reconstructed strands'), making this the load-bearing premise of the strand-level claim.
invented entities (1)
  • Strand Consistency (SC) metric no independent evidence
    purpose: Quantify topological accuracy of predicted strands by the highest fraction of a ground-truth strand's matched points contained in one predicted strand (Eq. 8).
    A new evaluation construct with no external validation. It penalizes fragmentation but not spurious connections, so a prediction merged into a single mega-strand would inflate SC. As a topology proxy it is one-sided and weak.

pith-pipeline@v1.3.0-alltime-deepseek · 9607 in / 22243 out tokens · 235758 ms · 2026-08-04T21:41:47.753465+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HairGS: Hair Strand Reconstruction based on 3D Gaussian Splatting." pith.science (2026). https://pith.science/paper/7D3FQLEF

@misc{pith2026250907774,
  author       = {Pith},
  title        = {Pith review of: HairGS: Hair Strand Reconstruction based on 3D Gaussian Splatting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7D3FQLEF}},
  note         = {Machine review of arXiv:2509.07774}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Human hair reconstruction is a challenging problem in computer vision, with growing importance for applications in virtual reality and digital human modeling. Recent advances in 3D Gaussians Splatting (3DGS) provide efficient and explicit scene representations that naturally align with the structure of hair strands. In this work, we extend the 3DGS framework to enable strand-level hair geometry reconstruction from multi-view images. Our multi-stage pipeline first reconstructs detailed hair geometry using a differentiable Gaussian rasterizer, then merges individual Gaussian segments into coherent strands through a novel merging scheme, and finally refines and grows the strands under photometric supervision. While existing methods typically evaluate reconstruction quality at the geometric level, they often neglect the connectivity and topology of hair strands. To address this, we propose a new evaluation metric that serves as a proxy for assessing topological accuracy in strand reconstruction. Extensive experiments on both synthetic and real-world datasets demonstrate that our method robustly handles a wide range of hairstyles and achieves efficient reconstruction, typically completing within one hour. The project page can be found at: https://yimin-pan.github.io/hair-gs/

Figures

Figures reproduced from arXiv: 2509.07774 by Matthias Nie{\ss}ner, Tobias Kirschstein, Yimin Pan.

Figure 2
Figure 2. Figure 2: An overview of the full pipeline. The process begins with image preprocessing and geometry reconstruction using 3D Gaussians. Hair strands are created by merging Gaus￾sians, followed by a refinement stage to optimize joint positions and the topology. whether the current number of points is sufficient or if further densification is required. 3.1 Stage I: Geometry Reconstruction The first stage of our pipeli… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the first-stage process. From an initial point cloud, 3D Gaussians are optimized using the differentiable rasterizer with the supervision of a combination of RGB, orientation, and mask losses (LRGB, Lθ , Lm). The densification process populates Gaussians in areas with missing geometry. between the precomputed mask M(x,y) and the rendered mask Mˆ(x,y): Lm = − 1 HW H ∑ y=1 W ∑ x=1 [PITH_FULL_IMA… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study. Full method gives best results. which evaluates the connectivity of points within hair strands. Specifically, the metric mea￾sures, for each Ground truth (GT) strand, the highest fraction of its points that are matched to points in a single predicted strand, and then averages this value over all strands. Let SG and SP denote the sets of ground truth and predicted strands, respectively. And … view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Wispy to Voluminous: Prior-free Multi-view Capture of Strand-level Facial Hair

    cs.GR 2026-06 unverdicted novelty 6.0

    Four-stage pipeline reconstructs explicit facial hair strands from multi-view images via 3D Gaussian optimization and curve tracing, claimed as first such method.

Reference graph

Works this paper leans on

36 extracted references · 23 canonical work pages · cited by 1 Pith paper · 3 internal anchors

  1. [1]

    Structure-from-motion using lines: Representation, triangulation, and bundle adjustment.Computer Vision and Image Understanding, 100(3):416–441, 2005

    Adrien Bartoli and Peter Sturm. Structure-from-motion using lines: Representation, triangulation, and bundle adjustment.Computer Vision and Image Understanding, 100(3):416–441, 2005. ISSN 1077-3142. doi: https://doi.org/10.1016/j.cviu.2005.06

    work page doi:10.1016/j.cviu.2005.06 2005

  2. [2]

    URLhttps://www.sciencedirect.com/science/article/pii/ S1077314205000846

    work page

  3. [3]

    Wide-baseline stereo matching with line segments

    Herbert Bay, Vittorio Ferrari, and Luc Van Gool. Wide-baseline stereo matching with line segments. volume 1, pages 329–336, 01 2005. doi: 10.1109/CVPR.2005.375

    work page doi:10.1109/cvpr.2005.375 2005

  4. [4]

    A morphable model for the synthesis of 3d faces

    V olker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. InProceedings of the 26th Annual Conference on Computer Graphics and Interactive PAN ET.AL.: HAIR STRAND RECONSTRUCTION BASED ON 3D GAUSSIAN SPLA TTING11 Techniques, SIGGRAPH ’99, page 187–194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. ISBN 0201485605. doi: 10.1145/...

    work page arXiv 1999

  5. [5]

    Single-view hair modeling using a hairstyle database.ACM Trans

    Liwen Hu, Chongyang Ma, Linjie Luo, and Hao Li. Single-view hair modeling using a hairstyle database.ACM Trans. Graph., 34(4), July 2015. ISSN 0730-0301. doi: 10.1145/2766931. URLhttps://doi.org/10.1145/2766931

    doi:10.1145/2766931 2015

  6. [6]

    Zhanghan Ke, Jiayu Sun, Kaican Li, Qiong Yan, and Rynson W.H. Lau. Modnet: Real-time trimap-free portrait matting via objective decomposition. InAAAI, 2022

    work page 2022

  7. [7]

    3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graph- ics, 42, 8 2023

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graph- ics, 42, 8 2023. ISSN 15577368. doi: 10.1145/3592433

    doi:10.1145/3592433 2023

  8. [8]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 12

    work page

  9. [9]

    Harold W. Kuhn. The hungarian method for the assignment problem.Naval Re- search Logistics (NRL), 52, 1955. URLhttps://api.semanticscholar. org/CorpusID:9426884

    work page 1955

  10. [10]

    Nersemble: Multi-view radiance field reconstruction of human heads.ACM Trans

    Tobias Kirschstein, Shenhan Qian, Simon Giebenhain, Tim Walter, and Matthias Nießner. Nersemble: Multi-view radiance field reconstruction of human heads.ACM Trans. Graph., 42(4), July 2023. ISSN 0730-0301. doi: 10.1145/3592455. URL https://doi.org/10.1145/3592455

    doi:10.1145/3592455 2023

  11. [11]

    Tony Lindeberg.Scale Invariant Feature Transform, volume 7. 05 2012. doi: 10.4249/ scholarpedia.10491

    work page 2012

  12. [12]

    Black, Hao Li, and Javier Romero

    Tianye Li, Timo Bolkart, Michael J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. InACM Transactions on Graphics, volume 36. Association for Computing Machinery, 11 2017. doi: 10.1145/3130800. 3130813

    doi:10.1145/3130800 2017

  13. [13]

    Mixture of volumetric primitives for efficient neural ren- dering, 2021

    Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural ren- dering, 2021. URLhttps://arxiv.org/abs/2103.01954

    Pith/arXiv arXiv 2021

  14. [14]

    Cdgnet: Class dis- tribution guided network for human parsing

    Kunliang Liu, Ouk Choi, Jianming Wang, and Wonjun Hwang. Cdgnet: Class dis- tribution guided network for human parsing. pages 4463–4472, 06 2022. doi: 10.1109/CVPR52688.2022.00443

    arXiv 2022

  15. [15]

    Paris, T

    Linjie Luo, Hao Li, S. Paris, T. Weise, Mark Pauly, and S. Rusinkiewicz. Multi-view hair capture using orientation fields. pages 1490–1497, 06 2012. ISBN 978-1-4673- 1226-4. doi: 10.1109/CVPR.2012.6247838. 12PAN ET.AL.: HAIR STRAND RECONSTRUCTION BASED ON 3D GAUSSIAN SPLA TTING

    work page arXiv 2012

  16. [16]

    Gaussianhair: Hair modeling and rendering with light-aware gaussians

    Haimin Luo, Min Ouyang, Zijun Zhao, Suyi Jiang, Longwen Zhang, Qixuan Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Gaussianhair: Hair modeling and rendering with light-aware gaussians. 2 2024. URLhttp://arxiv.org/abs/2402.10483

    Pith/arXiv arXiv 2024

  17. [17]

    Refinement of hair geometry by strand integration initial point cloud from lpmvs result of refinement strand integra- tion, 2023

    Ryota Maeda, Kenshi Takayama, and Takafumi Taketomi. Refinement of hair geometry by strand integration initial point cloud from lpmvs result of refinement strand integra- tion, 2023. URLhttps://github.com/elerac/strand_integration

    work page 2023

  18. [18]

    Wide-baseline hair capture using strand-based refinement

    Linjie Luo, Cha Zhang, Zhengyou Zhang, and Szymon Rusinkiewicz. Wide-baseline hair capture using strand-based refinement. pages 265–272, 06 2013. doi: 10.1109/ CVPR.2013.41

    work page 2013

  19. [19]

    Strand-accurate multi-view hair capture, 2019

    Giljoo Nam, Chenglei Wu, Min H Kim, and Yaser Sheikh. Strand-accurate multi-view hair capture, 2019. URLhttps://api.semanticscholar.org/CorpusID: 199016694

    work page 2019

  20. [20]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ra- mamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. 3 2020. URLhttp://arxiv.org/abs/2003.08934

    Pith/arXiv arXiv 2020

  21. [21]

    Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians.arXiv preprint arXiv:2312.02069, 2023

    Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Gieben- hain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians.arXiv preprint arXiv:2312.02069, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Capture of hair geometry from multiple images.ACM Transactions on Graphics, 23, 06 2004

    Sylvain Paris, Hector Briceño, and François Sillion. Capture of hair geometry from multiple images.ACM Transactions on Graphics, 23, 06 2004. doi: 10.1145/1015706. 1015784

    work page doi:10.1145/1015706 2004

  23. [23]

    Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys

    Johannes L. Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In Bastian Leibe, Jiri Matas, Nicu Sebe, and Max Welling, editors,Computer Vision – ECCV 2016, pages 501–518, Cham, 2016. Springer International Publishing. ISBN 978-3-319-46487-9

    work page 2016

  24. [24]

    Neural strands: Learning hair geometry and appearance from multi-view images

    Radu Alexandru Rosu, Shunsuke Saito, Ziyan Wang, Chenglei Wu, Sven Behnke, and Giljoo Nam. Neural strands: Learning hair geometry and appearance from multi-view images. 7 2022. URLhttp://arxiv.org/abs/2207.14067

    Pith/arXiv arXiv 2022

  25. [25]

    Dr.Hair: Reconstructing Scalp-Connected Hair Strands without Pre-training via Differentiable Rendering of Line Segments

    Yusuke Takimoto, Hikari Takehara, Hiroyuki Sato, Zihao Zhu, and Bo Zheng. Dr.hair: Reconstructing scalp-connected hair strands without pre-training via differentiable ren- dering of line segments, 2024. URLhttps://arxiv.org/abs/2403.17496

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Neural haircut: Prior-guided strand-based hair reconstruction

    Vanessa Sklyarova, Jenya Chelishev, Andreea Dogaru, Igor Medvedev, Victor Lempit- sky, and Egor Zakharov. Neural haircut: Prior-guided strand-based hair reconstruction. 6 2023. URLhttp://arxiv.org/abs/2306.05872

    Pith/arXiv arXiv 2023

  27. [27]

    NeuralHDHair: Automatic High-fidelity Hair Modeling from a Single Image Using Implicit Neural Representations

    Keyu Wu, Yifan Ye, Lingchen Yang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Neu- ralhdhair: Automatic high-fidelity hair modeling from a single image using implicit neural representations, 2022. URLhttps://arxiv.org/abs/2205.04175

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  28. [28]

    From multiview image curves to 3d drawings

    Anil Usumezbas, Ricardo Fabbri, and Benjamin Kimia. From multiview image curves to 3d drawings. volume 9908, pages 70–87, 10 2016. ISBN 978-3-319-46492-3. doi: 10.1007/978-3-319-46493-0_5

    work page doi:10.1007/978-3-319-46493-0_5 2016

  29. [29]

    Nofa: Nerf-based one-shot facial avatar reconstruction

    Wangbo Yu, Yanbo Fan, Yong Zhang, Xuan Wang, Fei Yin, Yunpeng Bai, Yan-Pei Cao, Ying Shan, Yang Wu, Zhongqian Sun, and Baoyuan Wu. Nofa: Nerf-based one-shot facial avatar reconstruction. pages 1–12, 07 2023. doi: 10.1145/3588432.3591555

    work page arXiv 2023

  30. [30]

    Monohair: High-fidelity hair modeling from a monocular video, 2024

    Keyu Wu, Lingchen Yang, Zhiyi Kuang, Yao Feng, Xutao Han, Yuefan Shen, Hongbo Fu, Kun Zhou, and Youyi Zheng. Monohair: High-fidelity hair modeling from a monocular video, 2024. URLhttps://arxiv.org/abs/2403.18356. PAN ET.AL.: HAIR STRAND RECONSTRUCTION BASED ON 3D GAUSSIAN SPLA TTING13

    Pith/arXiv arXiv 2024

  31. [31]

    Human hair reconstruction with strand-aligned 3d gaussians.ArXiv, September 2024

    Egor Zakharov, Vanessa Sklyarova, Michael J Black, Giljoo Nam, Justus Thies, and Otmar Hilliges. Human hair reconstruction with strand-aligned 3d gaussians.ArXiv, September 2024

    work page 2024

  32. [32]

    Hair meshes.ACM Transactions on Graphics, 28:1–7, 12 2009

    Cem Yuksel, Scott Schaefer, and John Keyser. Hair meshes.ACM Transactions on Graphics, 28:1–7, 12 2009. ISSN 15577368. doi: 10.1145/1618452.1618512

    work page arXiv 2009

  33. [33]

    HairNet: Single-View Hair Reconstruction using Convolutional Neural Networks

    Yi Zhou, Liwen Hu, Jun Xing, Weikai Chen, Han-Wei Kung, Xin Tong, and Hao Li. Hairnet: Single-view hair reconstruction using convolutional neural networks. 6 2018. URLhttp://arxiv.org/abs/1806.07467

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  34. [34]

    Meng Zhang and Y . Zheng. Hair-gan: Recovering 3d hair structure from a single image using generative adversarial networks.Visual Informatics, 3:102–112, 6 2019. ISSN 2468502X. doi: 10.1016/j.visinf.2019.06.001

    work page doi:10.1016/j.visinf.2019.06.001 2019

  35. [36]

    Groomcap: High-fidelity prior-free hair capture, 09 2024

    Yuxiao Zhou, Menglei Chai, Daoye Wang, Sebastian Winberg, Erroll Wood, Kri- pasindhu Sarkar, Markus Gross, and Thabo Beeler. Groomcap: High-fidelity prior-free hair capture, 09 2024

    work page 2024

  36. [2014]

    URLhttp://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv