acceptodds
Under review as a conference paper at ICLR 2027

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

11 views

Abstract

Realistic and editable animal fur reconstruction from multi-view images is challenging - fine-scale detail, self-occlusion and obfuscation, and, unlike human hair, there are no animal fur datasets, and fur usually covers most of the animal's body - with large inter/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 surface constrained gaussian "frosting" representation, along with part-based priors. We next show using a PCA-based decoder using knowledge from human hair strand data - allows us to alleviate the animal data scarcity, while allowing for faster optimization. FurE achieves a 10 strand training speedup over current SOTA dense per-strand optimization, retaining strand fidelity and generalizing across synthetic and more importantly, real-world sequences, with quantitative and quantitative validation despite this reduction in training time.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Strand-based fur reconstruction, multi-view 3D reconstruction, Gaussian splatting, PCA-based latent decoder, cross-domain (human-to-animal) hair priors.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Wf7sOpenReview ↗
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
AuthorsOpenReview ↗

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)D_\psi(z_i) = iDFT(\bar{S} + z_i^T X) 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.

AnimalLength InitμL\mu_LσL\sigma_LVarglobal\text{Var}_{\text{global}}Varloc\text{Var}_{\text{loc}}Varloc(dir)\text{Var}_{\text{loc}}(\text{dir})Varlocfirst(dir)\text{Var}_{\text{loc}}^{\text{first}}(\text{dir})
Pandav15.221.380.00040.0000250.04600.011
v25.231.380.00040.0000250.04900.014
Foxv13.801.150.000360.0000160.09400.017
v23.481.160.000410.0000160.09800.018
Catv11.600.780.000250.000010.0850.013
v21.620.480.000350.0000150.0880.015
WhiteTig.v13.701.600.00030.0000250.040.008
v23.821.680.000340.0000280.0430.01

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.

AnimalMethodIterationsμL\mu_LσL\sigma_LVarglobal\text{Var}_{\text{global}}Varloc\text{Var}_{\text{loc}}Varloc(dir)\text{Var}_{\text{loc}}(\text{dir})Varlocfirst(dir)\text{Var}_{\text{loc}}^{\text{first}}(\text{dir})
PandaNeuralFur22K5.231.380.00040.0000410.04300.081
FurE22K5.231.380.000410.0000230.04600.011
FoxNeuralFur22K3.001.180.000500.0000180.08600.019
FurE22K3.801.160.000350.0000160.09380.017
CatNeuralFur22K1.650.530.000600.0000850.0360.010
FurE22K1.610.780.000260.0000090.0850.013
WhiteTig.NeuralFur22K4.001.800.00030.0000180.050.013
FurE22K3.711.580.00030.0000240.040.009

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.

AnimalMethodμL\mu_LσL\sigma_LVarglobal\text{Var}_{\text{global}}Varloc\text{Var}_{\text{loc}}Varloc(dir)\text{Var}_{\text{loc}}(\text{dir})Varlocfirst(dir)\text{Var}_{\text{loc}}^{\text{first}}(\text{dir})
PandaNeuralFur(PCA-dec)5.251.400.00040.0000410.04220.082
FoxNeuralFur(PCA-dec)3.011.150.000360.0000160.09400.017
CatNeuralFur(PCA-dec)1.600.780.000250.000010.0850.013
WhiteTig.NeuralFur(PCA-dec)3.701.600.00030.0000250.040.008

Using local Frosting cues rather than part-level priors alone : All our reported results use the V1 variant from the local frosting cues.

AuthorsOpenReview ↗

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.

MethodPrecision (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 Length23.4036.2045.1019.1032.5043.4021.0334.2544.23
Using only shape Prior25.8039.1048.2020.4035.1046.9022.7836.9947.54
w/o Defurring21.8034.1042.8018.2030.8041.2019.8432.3741.98
Ours27.6241.7851.2021.6938.0050.5224.3039.8050.86

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.

Reviewer_jPPUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces FurE, an efficient framework for instance-specific 3D animal fur reconstruction from calibrated multi-view images. The method combines Gaussian Frosting-based defurring with a compact strand representation derived from a human-hair PCA prior. Instead of directly optimizing dense 3D strand geometry, FurE optimizes low-dimensional latent codes and decodes them into explicit, editable fur strands. Proposed method shows good performance on synthetic animal assets and a real-world bison sequence.

Critical strengths
  1. The paper addresses a practically relevant problem. Reconstructing explicit and editable animal fur without dedicated animal-fur training datasets is challenging and valuable for animal appearance reconstruction and 3D generation tasks.

  2. The proposed latent strand representation substantially reduces optimization time compared with dense per-strand reconstruction.

Critical weaknesses

Major comments:

  1. A central contribution is transferring a human-hair PCA representation to animal fur reconstruction. However, related ablations on human-hair PCA prior are missing.

  2. The proposed defurring stage relies on local Gaussian Frosting shell-width cues and part-level priors to estimate the underlying strand-root surface. However, the evaluation is primarily qualitative, and it remains unclear how closely the estimated defurred meshes match the actual body geometry.

Minor comments:

  1. The related work is not sufficiently comprehensive, as several recent advances in animal 3D reconstruction, animal fur and human hair reconstruction are missing.
Questions for authors
  1. Previous methods such as Neural Haircut and CGHair have already explored latent hair representations and efficient strand reconstruction. Can authors further clarify the novelty of proposed approaches?
3 comments
AuthorsOpenReview ↗

Response to Review

Related ablations on human-hair PCA prior are missing : We thank the reviewer for pointing out the missing ablation. We clarify that our PCA decoder is initialized with a pretrained human-hair prior; to isolate its impact, we compare this against a randomly initialized decoder on the ground-truth synthetic tiger dataset in the table below.

VariantPrecision (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)
Random Init18.1927.0236.7811.0322.1335.4013.7324.3336.08
Using Hair Prior27.6241.7851.2021.6938.0050.5224.3039.8050.86

Related work is not sufficiently comprehensive...: To the best of our knowledge, NeuralFur was the only published method addressing strand-based animal fur reconstruction from multi-view images at the time of submission. We welcome any specific references from the reviewer and will be happy to incorporate them into the updated manuscript.

Previous methods such as Neural Haircut and CGHair have already explored...: While Neural Haircut and CGHair target human scalp hair, FurE addresses the unique bottlenecks of full-body animal fur like lack of 3D ground truth strands, skin-depth ambiguity, and part-level variation. Unlike Neural Haircut, which relies on human-hair datasets and struggles with regional fur lengths (Table-3, Figure6 in the main paper), FurE adapts a lightweight human-hair PCA basis (PERM) to animals without any specific animal-fur training data via part level and thickness parameters.

Template-Free Shell-Based Defurring: Rather than assuming a known a fixed FLAME head scalp surface across different sequences like in Neural HairCut and CGHair, FurE combines local Gaussian Frosting thickness cues, 3D part segmentation, and bounded optimization to recover the hidden skin mesh and adapts to any quadruped species.

Thanks for the author feedback. The additional experiments on human-hair PCA prior are useful. However, the authors did not answer my major cmments 2. I would like to maintain my initial rating.

AuthorsOpenReview ↗

Response to Reviewer jPPU

We thank you for the response on our rebuttal. Below, we address your comment in detail.

The proposed defurring stage relies on ..... actual body geometry.

Absence of Ground-Truth Skin Meshes: A direct quantitative comparison of the estimated defurred mesh is not possible because acquiring real-world 3D ground truth of the bare skin beneath live animal fur is physically impossible and inhumane. Furthermore, publicly available synthetic datasets provide only posed multi-view images without an accompanying ground-truth bare skin mesh, hence we are unable to perform any direct mesh-to-mesh evaluation.

However , we provide the quantitative analysis of the mesh geometry from the artist created synthetic tiger in the table below

MethodChamfer Distance (↓\downarrow)
NeuralFur5.82 * 1e-4
Ours4.68 * 1e-5

Note that to fit within the rebuttal time window, we ran NeuralFur’s preprocessing with reduced optimization steps; we will include a thorough quantitative analysis of the full defurring procedure on the synthetic tiger in the revised manuscript. Importantly, even if NeuralFur matches our accuracy at full optimization steps, our preprocessing remains ~10× faster.

Reviewer_jQwKOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents FurE, a method for instance-specific 3D animal fur reconstruction from calibrated multi-view images without requiring animal-fur training data. The method first estimates a defurred body surface using Gaussian Frosting thickness cues and part-level priors. It then reconstructs explicit fur strands using a root-conditioned latent representation and a PCA decoder initialized from human hair data. The reconstructed strands are optimized through differentiable rendering with strand-aligned Gaussians.

The method is evaluated on synthetic animals from Artemis and a real-world bison sequence. Compared to NeuralFur, FurE reduces strand training time from 10.5 hours to 52 minutes while achieving comparable rendering quality.

Critical strengths

[1] Relevant and challenging problem: Reconstructing explicit animal fur from multi-view images is a challenging problem due to fine-scale geometry, self-occlusion, and the lack of animal-fur training datasets. It is also a worthwhile problem to address, particularly for creating editable 3D animal assets.

[2] Efficient strand reconstruction: FurE substantially reduces strand optimization time compared to NeuralFur while maintaining comparable rendering quality. Since instance-specific reconstruction requires optimization for each animal, reducing the training time from 10.5 hours to 52 minutes is a useful improvement. The method also retains explicit, editable fur strands.

[3] Reconstruction without animal-fur training data: The use of a human-hair PCA representation for animal fur is an interesting approach to handling the lack of animal-fur datasets. The method produces reasonable reconstruction results across several synthetic animals and demonstrates its applicability to a real-world bison sequence.

Critical weaknesses

[1] Insufficient analysis of efficiency improvement: The paper reports around a 10× speedup over NeuralFur, but the two methods use different components, including NeuS vs. NeuS2 and SMAL fitting vs. 3D part segmentation. These differences may account for part of the preprocessing speedup.

For strand training, the contribution of the PCA representation is not clearly isolated. A comparison under the same strand count and computational budget, along with a direct strand-coordinate prediction baseline, would help clarify where the speedup comes from.

[2] Limited validation of real-world reconstruction and generalization: The paper reports that NeuralFur fails to reconstruct the underlying geometry of the real-world bison. However, since the two methods use NeuS2 and NeuS respectively, it is unclear whether this difference comes from the proposed method or the reconstruction backbone.

The evaluation mainly uses a small set of synthetic animals from Artemis with 36 input views. These controlled conditions may favor accurate geometry reconstruction, but the paper does not examine performance with fewer or less uniformly distributed views. The real-world evaluation is also limited to a bison sequence.

The paper also claims applicability to non-quadruped mammals, but all evaluated animals are quadrupeds, and no quantitative part segmentation evaluation is provided. The broader generalization claims are therefore not sufficiently supported.

[Minor weaknesses]

  • Implementation details: The use of ALIGN-Parts, including its training setup and manual verification, is not sufficiently described. The VLM model and prompts used for fur length estimation in V2 are also unclear.

  • Presentation: The plots and labels in Figure 4 are difficult to read, and the differences between the defurred surfaces in Figure 5 are not clearly visible. There are also formatting issues around lines 184 and 209.

Questions for authors

[1] Efficiency comparison: Can the authors clarify how much of the preprocessing speedup comes from using NeuS2 and removing SMAL fitting? For strand training, although the appendix states that the same hyperparameters as NeuralFur were used, the comparison in terms of total strand count, trainable parameters, and computational cost is unclear. Could the authors provide these details and a controlled comparison between the PCA decoder and direct strand-coordinate prediction using the same root-conditioned encoder?

[2] Real-world reconstruction: It is unclear whether NeuralFur fails on the bison sequence due to its NeuS-based initial surface reconstruction or the subsequent defurring and strand reconstruction stages. Would NeuralFur still fail if both methods used the same initial geometry reconstructed by NeuS2? Also, how were the camera intrinsics and extrinsics obtained for the real-world bison sequence, and how many views were used?

[3] Contribution of VLM priors and frosting-based initialization: FurE introduces both Frosting-based length estimation (V1) and a variant that combines it with VLM priors (V2). Which variant was used in the main experiments, and what is the quantitative performance difference between V1 and V2? NeuralFur also uses VLM-based fur length and thickness priors. Did FurE use the same VLM settings as NeuralFur? What are the benefits of Frosting-based initialization compared to the original VLM-based priors?

[4] 3D part segmentation: Was ALIGN-Parts used directly with pretrained weights, or was additional training or fine-tuning required? How much manual correction was needed during part annotation? Given the calibrated multi-view images and reconstructed mesh, why was ALIGN-Parts chosen over simpler approaches that obtain semantic part masks from 2D foundation models and fuse them into 3D? How dependent is the reconstruction on the accuracy of the part labels?

[5] Implementation details: Which VLM and prompts were used for part-specific fur length estimation in V2? How were the absolute fur lengths estimated in centimeters, and how sensitive is the final reconstruction to errors in these estimates?

3 comments
AuthorsOpenReview ↗

Response to Review

Insufficient analysis of efficiency improvement : We would like to clarify that our core claim of a 10x speedup specifically refers to the Stage 2 strand training phase, which is evaluated completely independently of the Stage 1 preprocessing components. Nevertheless, as reported in Table 1 of the main paper (measured on the same NVIDIA A5000 GPU across 36 views), we explicitly separate the pipeline into two distinct stages and report the time for both.

For further clarification we report timings for each step in Stage1

  • Mesh generation : 1 hour.
  • Mesh extraction : 49-183 seconds.
  • Mesh segmentation : 1-2 sec.

We also clarify that in all our comparisons against the baselines we maintain strict parity across the optimization setup: we use the exact same strand count (15K sampled strands per iteration) on the same hardware (NVIDIA A5000 GPU) and same number of views (36).

Limited validation of real-world reconstruction and generalization ... We thank the reviewer for this constructive suggestion . For validating the failure of NeuralFur on Bison we use the MLP decoder from NeuralFur combined with our preprocessing pipeline to reconstruct the strands. The quantitative results clearly show the degradation in quality for the generated fur as compared to ours. We use COLMAP to predict the camera parameters for the real world example and we use 200 views for training.

MethodμL\mu_LσL\sigma_LVarglobal\text{Var}_{\text{global}}Varloc\text{Var}_{\text{loc}}Varloc(dir)\text{Var}_{\text{loc}}(\text{dir})Varlocfirst(dir)\text{Var}_{\text{loc}}^{\text{first}}(\text{dir})PSNRSSIMLPIPSTime
NF decoder7.002.035e-43.8e-40.060.1343.110.38800.489710hrs
Our decoder11.101.602.6e-43.5e-40.030.00844.100.41890.488052min

Contribution of VLM priors and frosting-based initialization : We clarify that we use V1 as our default setting, hence our results demonstrate a purely geometric, non-VLM based strand length initialization . We compare the quantitative performance against the V1 variant in this table where we use V1/V2 variants to get the final length.

AnimalLength InitμL\mu_LσL\sigma_LVarglobal\text{Var}_{\text{global}}Varloc\text{Var}_{\text{loc}}Varloc(dir)\text{Var}_{\text{loc}}(\text{dir})Varlocfirst(dir)\text{Var}_{\text{loc}}^{\text{first}}(\text{dir})
Pandav15.221.380.00040.0000250.04600.011
v25.231.380.00040.0000250.04900.014
Foxv13.801.150.000360.0000160.09400.017
v23.481.160.000410.0000160.09800.018
Catv11.600.780.000250.000010.0850.013
v21.620.480.000350.0000150.0880.015
WhiteTig.v13.701.600.00030.0000250.040.008
v23.821.680.000340.0000280.0430.01

Did FurE use the same VLM settings as NeuralFur ? No, we calculate the fur length and thickness prior as explained in Appendix A1.

What are the benefits of Frosting-based initialization compared to the original VLM-based priors? VLM priors output a single value per body part based solely on species appearance and ignoring the actual 3D volume of the captured subject. In contrast, Gaussian Frosting is optimized directly on the target animal's calibrated multi-view images and captures view-consistent, instance-specific 3D volumetric thickness and length.. Figure 11 in the Appendix visualizes how our Frosting-based initialization adapts to instance-specific regional variations (blue/red offsets) that uniform VLM priors miss.

Was ALIGN-Parts used directly with pretrained weights, or was additional training or fine-tuning required? Yes we use it directly without any finetuning .

Given the calibrated multi-view images and reconstructed mesh, why was ALIGN-Parts chosen over simpler approaches that obtain semantic part masks from 2D foundation models and fuse them into 3D? How dependent is the reconstruction on the accuracy of the part labels? Animal fur reconstructions involve fine-scale anatomical structures (see Figure 3) that suffer from severe multi-view self-occlusion. Lifting 2D foundation model masks (SAM/Grounded-SAM) into 3D via multi-view projection frequently suffers from cross-view label inconsistency. While accurate labels are important for non-furred regions our pipeline is robust to errors between fur bearing parts because of local 3D frosting estimates.

Implementation details: We use the same VLM prompts as NeuralFur. We explain the process of converting fur length to centimeters in the Appendix A1 Section : Metric scale and length statistics. As shown in the above table comparing V1 and V2 , our method is robust to errors in length estimates.

Thank you for the detailed rebuttal and additional experiments. It clarify several of my concerns, particularly regarding efficiency. However, the individual contributions of the proposed components and the generalization across diverse observation conditions remain insufficiently validated, so I maintain my initial rating.

AuthorsOpenReview ↗

Response to Reviewer jQwK

Thank you for taking the time to review our rebuttal and for acknowledging our efficiency clarifications. To validate both individual component contributions and generalization across diverse animal species, we ablated each core module and evaluated our method across four diverse animal sequences (Panda, Fox, Cat, and White Tiger), alongside a noisy real-world capture (Bison).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.