PaperlayerView exact revision on arXiv

Source-linked AI summary

FurE: Efficient Instance-Specific 3D Fur Reconstruction without Animal-Fur Datasets

Srinjay Sarkar, Prakhar Kaushik, Soumava Paul, Alan Yuille

arXiv:2609.35770v1cs.CVcs.AIcs.GR

TL;DR

Animal-fur reconstruction from images is difficult because strands are occluded, the underlying root surface is ambiguous, fur varies widely, and animal-fur training data are lacking. FurE combines shell-based defurring with human-hair-informed compact strand learning. On Artemis, it reports 10× faster strand training than NeuralFur with comparable rendering quality, and reconstructs fur from a noisy real-world bison sequence on which NeuralFur fails.

  • Problem

    Animal fur is difficult to reconstruct as explicit strands because fine detail is occluded, the visible envelope obscures the root surface, fur varies widely, and animal-fur training data are absent.

    Source: §1 · pp. 1–2

  • Method

    FurE combines Gaussian Frosting cues and part-aware priors for defurring with root-conditioned latent codes decoded into explicit strands by a human-hair PCA prior.

    Source: §1 · p. 2

  • Results

    On Artemis, FurE achieves a 10× strand-training speedup over NeuralFur with comparable rendering quality, and reconstructs a noisy real-world bison sequence on which NeuralFur fails.

    Source: §5 · p. 9

  • Takeaways & Limitations

    FurE retains explicit strands for downstream editing, rendering, and simulation while avoiding a separate animal-fur training dataset.

    Source: §1 · p. 2, §5 · p. 9

    Evaluation lacks ground-truth cases with strong within-part variation, and FurE uses a single strand layer on a root surface estimated before strand optimization.

    Source: §5 · p. 10

Abstract

from arXiv · show

Realistic and editable animal fur reconstruction from multi-view images is challenging due to fine-scale detail, self-occlusion and obfuscation, and, unlike human hair, the lack of animal-fur datasets. Fur usually covers most of an animal's body, with large inter-species and intra-species variability. We present FurE, an efficient strand-based animal fur reconstruction method that recovers a per-strand, editable groom by optimizing a root-conditioned latent field, decoded into strand geometry via a PCA-based decoder. We reconstruct a defurred animal body using local fur-thickness cues from a surface-constrained Gaussian Frosting representation together with part-based priors. We further show that a PCA-based decoder learned from human-hair strand data can alleviate animal-data scarcity while enabling substantially faster optimization. FurE achieves a 10x speedup in strand training over current SOTA dense per-strand optimization while retaining strand fidelity and generalizing across synthetic and real-world sequences, with quantitative and qualitative validation despite the reduction in training time.

1 INTRODUCTION

FurE addresses the occlusion, root-surface ambiguity, and variability that make explicit animal-fur reconstruction difficult. It combines shell-based defurring with compact, human-hair-informed strand learning, reporting faster training and reconstruction from noisy real-world multiview images.

1 INTRODUCTION

  • Animal-fur reconstruction is difficult because fine strands are self-occluded, the visible surface is an outer furry envelope rather than the root surface, and fur varies across species and body parts.The fur also covers most of the animal’s visible body.Source: §1 · p. 1
  • A pretrained PCA decoder maps compact, root-conditioned codes on the defurred body to explicit strand geometry, reducing per-scene learning cost relative to dense strand optimization.The decoder uses human-hair data, avoiding a separate animal-fur training dataset.Source: §1 · p. 2
  • A 10× strand-training speedup with comparable or better quantitative results is reported on Artemis, alongside reconstruction from noisy real-world multiview images where NeuralFur struggles.The method retains explicit, editable strands.Source: §1 · p. 2
  • FurE estimates the hidden strand-root surface by combining local Gaussian Frosting cues with part-aware priors, without SMAL fitting.It treats defurring as local shell estimation, using fur-depth cues from Gaussian Frosting.Source: §1 · p. 2
  • FurE demonstrates instance-specific strand-based animal-fur reconstruction from noisy real-world multiview images.The authors describe this as the first such demonstration to their knowledge.Source: §1 · p. 2

2 RELATED WORK

Prior animal-reconstruction methods recover body shape but not explicit fur strands, while prior strand methods focus on hair or incur costly dense optimization. FurE combines local defurring cues with compact latent strand learning.

Animal reconstruction

  • Unlike body-shape methods that do not recover explicit fur, FurE uses local Gaussian Frosting cues and part-aware priors for the hidden root surface, with direct part segmentation removing SMAL fitting.NeuralFur instead relies on template fitting and part-level semantic estimates, while dense strand optimization remains computationally expensive.Source: §2 · p. 2

Strand-based and compact hair reconstruction

  • FurE transfers a human-hair PCA prior to animal fur and optimizes root-conditioned latent codes rather than dense strand geometry, reducing per-scene cost without animal-fur training data.It retains explicit, editable strands.Source: §2 · p. 3

Figure 2

  • Figure 2 presents a two-stage workflow: Gaussian Frosting-based mesh defurring followed by PCA-based strand decoding, with an optional texture generator.The method is optimized end to end from multiview images in a 3DGS-compatible framework.Source: Figure 2 · p. 3

3 METHOD

FurE estimates a defurred body and initializes strand attributes from local shell cues and part-aware guidance, then learns editable strand geometry in a compact latent space. Differentiable rendering of strand-aligned Gaussians lets multiview losses refine the reconstruction.

3.1 DEFURRING AND FUR INITIALIZATION

  • For defurring, FurE moves the reconstructed outer mesh inward using local Frosting thickness cues, part-level guidance, smoothness, and geometry-aware bounds, while initializing strand lengths separately from root displacement.V1 uses the 75th-percentile shell width per part multiplied by a fixed part-specific factor; optional V2 blends this estimate with a VLM or supplied reference length.Source: §3 · p. 3, §3.1 · p. 4

3.2 CODEC-BASED STRAND LEARNING

  • A root-conditioned encoder predicts compact strand coefficients from root position, part label, local thickness, and target length; a human-hair PCA decoder maps them to explicit, editable animal-fur strands.The decoder is fine-tuned on the target animal’s multiview images, without a separate animal-fur training dataset.Source: §3 · p. 3, §3.2 · p. 5

3.3 RENDERING AND OPTIMIZATION

  • FurE represents decoded strands with cylindrical Gaussians and optimizes the latent field using multiview losses for appearance, masks, orientation, envelope fit, and body-penetration control.Each Gaussian follows a strand segment’s length and local tangent; the losses include RGB, silhouette, orientation, chamfer, SDF, and fur-mask terms.Source: §3.3 · p. 5

4 EXPERIMENTS

FurE is evaluated on synthetic animal-fur sequences and a real-world bison sequence, with comparisons spanning reconstruction quality, efficiency, and design choices. The results report competitive strand fidelity at substantially lower training cost, with further visualizations across animal subjects.

4 EXPERIMENTS

  • Five synthetic fur styles are evaluated using 36 training frames per sequence, with remaining frames reserved for novel-view rendering; a real-world bison sequence tests generalization.Comparisons include surface-reconstruction and strand-based baselines.Source: §4 · p. 5, Table 1 · p. 5
  • On a synthetic tiger with ground-truth strands, the quantitative protocol measures strand precision, recall, and F-score; four Artemis scenes use unsupervised geometry-consistency metrics.The latter assess length, direction, curvature, and proximity to the surface without ground-truth strands; rendered strand-aligned Gaussians are also assessed using PSNR, LPIPS, and SSIM.Source: §4 · pp. 5–7, Table 3 · p. 8, Table 4 · p. 8
  • FurE largely outperforms NeuralFur and Gaussian Haircut on rendering quality while training 10× faster; FurE and NeuralFur yield coherent strands, unlike coarse surface methods and artifact-prone Gaussian Haircut.The rendering comparison covers four scenes; the runtime comparison uses 36 views and an A5000 GPU.Source: Table 1 · p. 5, Figure 6 · p. 6, §4 · p. 7
  • Surface methods lack strand-level detail, while Gaussian Haircut produces inconsistent lengths and artifacts; FurE and NeuralFur recover accurate, coherent strand geometry.The comparison reports FurE as 10× faster.Source: Figure 6 · p. 6, §4 · p. 7
  • Frosting supplies a local, view-consistent shell-width cue, a lower bound on strand arc length that FurE calibrates with part-level priors; defurring results are similar to NeuralFur at significantly higher speed.The cited comparison visually validates the defurring approach and its fur-length and width calculations.Source: §4 · p. 7
  • Fixed strand lengths across semantic regions distort body and belly fur; removing the defurred mesh looks similar but yields inconsistent direction and curvature.These comparisons assess strand-length parameterization and the defurred mesh within the strand-generation pipeline.Source: Figure 8 · p. 7, §4 · p. 8
  • Reconstructed strands can be imported into Blender and Unreal Engine, and strong-wind simulation shows physically coherent motion.The wind response is demonstrated in supplementary video.Source: Figure 9 · p. 8, §4 · p. 9

4.1 MORE VISUALIZATIONS

  • Novel-view renderings of panda, white tiger, fox, and cat show instance-specific fur characteristics and high-fidelity geometric detail across viewpoints without dense per-strand optimization.Figure 10 provides additional renderings across these animal subjects.Source: §4.1 · p. 9, Figure 10 · p. 9

5 CONCLUSION

FurE reconstructs explicit, editable animal fur from calibrated multiview images using a human-hair PCA basis, Gaussian Frosting cues, and part-aware priors. On Artemis it trains strands in under an hour with a 10× speedup over NeuralFur and comparable rendering quality, while reconstructing a noisy real-world bison sequence on which NeuralFur fails.

5 CONCLUSION

  • FurE uses a human-hair PCA basis, local Gaussian Frosting cues, and part-aware priors to reconstruct explicit, editable animal fur without a separate animal-fur training dataset.Direct part segmentation removes the need for SMAL fitting.Source: §5 · p. 9
  • On Artemis, strand training takes under one hour, giving a 10× speedup over NeuralFur with comparable rendering quality; FurE also reconstructs a noisy bison sequence on which NeuralFur fails.The resulting explicit strands remain available for downstream editing, rendering, and simulation.Source: §5 · p. 9
  • The evaluation lacks ground-truth cases with strong within-part variation, and FurE models a single strand layer on one root surface estimated before strand optimization.The authors identify multilayer coats and joint root-placement and strand-geometry optimization as next steps.Source: §5 · p. 10

A.1 IMPLEMENTATION DETAILS FOR FROSTING-CALIBRATED DEFURRING

Frosting shell widths serve as relative local evidence for fur thickness, while part-level references calibrate a plausible root surface and strand lengths. The implementation keeps defurring and length initialization distinct and uses separate root-sampling and strand-width settings.

A.1 IMPLEMENTATION DETAILS FOR FROSTING-CALIBRATED DEFURRING

  • Frosting widths are transferred to the visible mesh as relative local cues, then calibrated with part-aware references rather than treated as direct skin-depth or strand-length measurements.Figure 11 contrasts raw local thickness cues with the final calibrated lengths, which combine them with shape and part priors.Source: §A.1 · p. 11, Figure 11 · p. 13
  • Defurring balances supported local thickness evidence, part-level guidance, and mesh-edge smoothing while bounding inward displacement and fixing protected non-fur regions.Unsupported vertices have zero correspondence support and their local term is omitted; missing support is not interpreted as absence of fur.Source: §A.1 · pp. 11–12
  • Part calibration caps shell widths at the part’s 95th percentile, blends the 75th-percentile statistic with a coarse thickness reference, and caps the result at the 95th percentile.When a reference is missing, the 75th-percentile statistic is used; an explicitly supplied zero remains zero.Source: §A.1 · p. 11
  • Length initialization uses raw unrestricted shell widths and requires metric scale for centimeter-valued estimates; V1 uses fixed part-specific multipliers, while V2 can blend a reference length.These variants change strand-length initialization, not the defurring objective; non-fur parts receive zero length.Source: §A.1 · p. 12
  • Root sampling assigns independent base densities by part and uses part labels and local Frosting cues, while sampled roots carry local frames, initialized lengths, and a normalized shell-width feature.The shell-width feature is distinct from correspondence support and root displacement.Source: §A.1 · p. 12

A.2 CODEC STRAND LEARNING DETAILS

FurE learns strands through a compact root-conditioned code decoded into normalized curves, rather than optimizing dense strand control points. Root and part information, density, and initialized length scale the decoded geometry.

A.2 CODEC STRAND LEARNING DETAILS

  • A root-conditioned encoder predicts a compact latent code for each sampled root, and a decoder maps it to a normalized strand curve.This replaces dense per-strand control-point optimization with compact strand codes.Source: §A.2 · p. 13
  • The root tangent frame, part label, Frosting density, and initialized strand length scale the decoded curve into world space.Source: §A.2 · p. 13
  • Training uses a continuous latent texture, 15K strands per iteration for 2,500 iterations, and fine-tunes the PCA-based decoder with NeuralFur’s hyperparameters.Source: §A.2 · p. 13

A.3 APPLICATIONS

FurE’s reconstructed strands can be imported into standard 3D game engines and used in downstream physics simulations. Wind simulation produces realistic, coherent motion.

A.3 APPLICATIONS

  • Reconstructed strands can be imported into commonly used 3D game engines, including Blender and Unreal Engine.Source: §A.3 · p. 13
  • The strands’ structural plausibility supports downstream physics simulations after import.Source: §A.3 · p. 13
  • Under simulated wind dynamics, the imported strands produce realistic and coherent motion.Source: §A.3 · p. 13

A.4 LIMITATIONS AND FUTURE WORK

FurE currently requires calibrated multiview images of a static animal, and dense capture can be difficult for living animals. The authors identify dynamic reconstruction and stronger geometric priors as future needs because severe occlusion can degrade shell-thickness and fur-length estimates.

A.4 LIMITATIONS AND FUTURE WORK

  • The framework requires calibrated multiview images of a static animal, making extension to dynamic motion a stated next step.Source: §A.4 · p. 13
  • Severe input-view or fur self-occlusion can degrade shell-thickness and fur-length estimates, motivating stronger geometric priors.Source: §A.4 · p. 13
  • The additional real-world examples include explicit editable strands reconstructed from noisy bison multiview images, with associated data released.Source: Figure 13 · p. 14

A.5 REAL WORLD FUR RECONSTRUCTION

On noisy real-world bison multiview images, FurE reconstructs strand-based geometry, while NeuralFur fails to recover the underlying surface; the authors also release the multiview data.

A.5 REAL WORLD FUR RECONSTRUCTION

  • FurE reconstructs strand-based geometry from noisy real-world multiview images in an instance-specific bison sequence.The authors describe this as a demonstration of generalizability and robustness.Source: §A.5 · p. 14
  • NeuralFur fails to capture the bison’s underlying geometric surface, preventing coherent and plausible fur geometry.The authors attribute the fur-geometry failure to surface degradation.Source: §A.5 · p. 14
  • The authors release the associated multiview data to encourage research on real-world fur reconstruction.Source: §A.5 · p. 14
Loading 2609.35770v1…