- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes FurE, a method for reconstructing explicit, editable animal fur strands from calibrated multiview images. It estimates an underlying defurred body surface using Gaussian Frosting thickness cues and body-part priors, then optimizes root-conditioned latent codes decoded into strands using a PCA representation derived from human hair. The method reports substantially faster reconstruction than NeuralFur while maintaining similar rendering quality.
Critical strengths
S1. The reported speedup is large: training takes 52 minutes instead of 10.5 hours, while still producing editable strands.
S2. The reported quality is maintained despite the speedup. Novel-view rendering metrics are nearly identical to NeuralFur, and the ground-truth strand F-score is slightly better on the synthetic tiger.
S3. Using a lightweight PCA decoder makes sense to me as a practical design choice for reducing optimization complexity.
S4. The figures are simple and easy to understand.
S5. The real-world bison reconstruction is a useful proof of concept beyond the synthetic examples.
Critical weaknesses
W1. NeuralFur already uses a root-conditioned network, low-dimensional codes, and a human-hair prior. PERM already introduced PCA-based strand representations. So, presenting the move to low-dimensional learning as a new contribution overstates the novelty. The main differences appear to be the choice of decoder and the defurring procedure, but the benefits of these changes are not sufficiently isolated experimentally.
W2. Figure 8 and Table 3 include some ablations, but they do not isolate the contributions of the proposed components to efficiency and reconstruction quality. In particular, the paper does not compare replacing only NeuralFur's decoder with PCA or removing only the local Frosting cues. Such experiments would help establish which changes are actually responsible for the improvements.
W3. The reported speedup is substantial, but the runtime comparison is not sufficiently explained. FurE uses 2,500 iterations, while NeuralFur's public example uses 20,000, and it is unclear what settings were used for the reported baseline. Without matched iteration budgets or quality-versus-runtime comparisons, it is difficult to determine whether the speedup comes from cheaper iterations or fewer optimization steps.
W4. The geometric accuracy of the reconstruction is not thoroughly validated. Ground-truth strand accuracy is only evaluated on one synthetic tiger, while the remaining geometric metrics mainly measure consistency rather than correctness. The authors could extend their existing ground-truth evaluation to the ablated variants and evaluate defurring accuracy against the underlying mesh of the tiger. Evaluating more varied fur geometries, particularly local variations within body parts, would also help establish the benefit of Frosting-based cues over coarse part-level priors.
W5. The ground truth evaluation only considers a single example. While this follows the protocol of the NeuralFur baseline, there are more synthetic datasets out there that could feasible be adapted to the pipeline and evaluated on, which would significantly strengthen the evaluation of the method. For example https://huggingface.co/datasets/Chunyi99/AnimalLift
Questions for authors
Q1. Could the authors clarify the experimental settings, particularly the PCA dimension, the details of decoder fine-tuning, V1/V2 choice, and whether the priors are manually adjusted for each instance?
Q2. How many strand-optimization iterations were used for NeuralFur in Table 1? Can the authors compare both methods at matched iteration counts or matched runtime and report the quality–runtime trade-off?
Q3. Can the authors isolate the effect of replacing NeuralFur’s decoder with PCA and the effect of using local Frosting cues rather than part-level priors alone, including both quality and runtime?
Q4. Could the authors extend the ground-truth strand evaluation to the ablated variants and quantitatively evaluate the defurred mesh against ground-truth geometry? It would also be interesting to see results on synthetic animals with local variations in fur length or density within the same body part.
7 comments
Response to Review (Part-2)
Could the authors extend the ground-truth strand evaluation to the ablated variants and quantitatively evaluate the defurred mesh against ground-truth geometry? It would also be interesting to see results on synthetic animals with local variations in fur length or density within the same body part. We provide the ground-truth strand evaluation to the ablated variants below. Our method uses multi-view posed images without any GT mesh , hence we cannot quantitatively compare the quality of our generated mesh.
| Method | Precision (2/20) | Precision (3/30) | Precision (4/40) | Recall (2/20) | Recall (3/30) | Recall (4/40) | F-score (2/20) | F-score (3/30) | F-score (4/40) |
|---|---|---|---|---|---|---|---|---|---|
| Fixed Length | 23.40 | 36.20 | 45.10 | 19.10 | 32.50 | 43.40 | 21.03 | 34.25 | 44.23 |
| Using only shape Prior | 25.80 | 39.10 | 48.20 | 20.40 | 35.10 | 46.90 | 22.78 | 36.99 | 47.54 |
| w/o Defurring | 21.80 | 34.10 | 42.80 | 18.20 | 30.80 | 41.20 | 19.84 | 32.37 | 41.98 |
| Ours | 27.62 | 41.78 | 51.20 | 21.69 | 38.00 | 50.52 | 24.30 | 39.80 | 50.86 |
The ground truth evaluation only considers a single example. While this follows the protocol of the NeuralFur baseline, there are more synthetic datasets out there that could feasible be adapted to the pipeline and evaluated on, which would significantly strengthen the evaluation of the method. For example https://huggingface.co/datasets/Chunyi99/AnimalLift
Acquiring real-world 3D ground truth for both individual fur strands and the underlying bare skin of live animals is neither feasible nor humane. As a result, evaluation relies on artist-groomed 3D assets, which are scarce and expensive due to the labor-intensive nature of manual strand modeling. Following NeuralFur protocol, we evaluate on the synthetic tiger as it is a readily available high-quality asset publicly available with GT fur strands . Notably, overcoming this exact scarcity of labor-intensive 3D assets is a primary motivation for our method.
Thank you to the authors for their detailed response and additional experiments, and apologies for my late reply. I appreciate the effort to address my concerns, particularly the new ground-truth ablations. However, several points remain unresolved:
- W1: The methodological novelty beyond NeuralFur and PERM remains unclear.
- W3: The matched-iteration results lack runtime measurements, leaving the source of the speedup unclear.
- W4: The additional strand ablations are helpful, but a direct quantitative comparison of the estimated defurred body mesh against the ground-truth body mesh is still missing.
- W5: Ground-truth evaluation remains limited to one animal, without addressing the possibility of using additional synthetic datasets.
While the rebuttal strengthens the evaluation, I believe the main concerns regarding novelty, efficiency, and geometric validation remain insufficiently addressed. I therefore maintain my weak reject rating.
Response to Review
Could the authors clarify the experimental settings, particularly the PCA dimension ..... for each instance? We use a 64-dimensional frequency-domain PCA coefficient vector, partitioned into 10 low-frequency coefficients that capture coarse strand shape and 54 higher-order coefficients that capture local curvature.The PCA decoder Dψ(zi)=iDFT(Sˉ+ziTX) is a lightweight linear basis projection in the Discrete Fourier Transform (DFT) domain, where S is the generated strand vector and X is the pretrained human-hair principal component matrix (initialized from PERM), followed by an inverse DFT , our root-frame TBN transformation and part-length scaling. We use the same prior for every sample.
We compare the quantitative performance against the V1 variant in this table where we use V1/V2 variants to get the final length.
How many strand-optimization iterations were used for NeuralFur in Table 1? Can the authors compare both methods at matched iteration counts or matched runtime and report the quality–runtime trade-off? For all reported results of NeuralFur we use 22K iterations. We run FurE for 22K iterations and report the results in the following table. We observe very little quantitative difference in our results as compared to 2.5K iterations suggesting extremely fast convergence for our method.
Effect of replacing NeuralFur’s decoder with PCA and the effect of using local Frosting cues rather than part-level priors alone We show that replacing the NeuralFur decoder with ours generates strands of similar quality but significantly cuts down on generation time from 10.5 hours to 52 mins demonstrating our strand decoder is compatible with other preprocessing methods as well.
Using local Frosting cues rather than part-level priors alone : All our reported results use the V1 variant from the local frosting cues.