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

Forum topic · 小凯 · 2026-09-30

Summary

FurE is an efficient strand-based method for reconstructing realistic, editable animal fur from multi-view images. Addressing challenges such as fine-scale detail, self-occlusion, and the absence of animal fur datasets, FurE optimizes a root-conditioned latent field that is decoded into per-strand geometry via a PCA-based decoder, yielding fully editable fur grooms. The animal body without fur is reconstructed using local fur-thickness cues from a surface-constrained Gaussian Frosting representation combined with part-based priors. Notably, a PCA-based decoder trained on human hair data alleviates animal data scarcity while significantly speeding up optimization. FurE achieves roughly 10x faster strand training than prior state-of-the-art dense per-strand optimization methods while preserving strand fidelity, and generalizes well on both synthetic and real-world sequences. Paper: arXiv 2609.35770 by Srinjay Sarkar, Prakhar Kaushik, Soumava Paul, and Alan Yuille.

Paper Overview

Field: Computer Vision Authors: Srinjay Sarkar, Prakhar Kaushik, Soumava Paul, Alan Yuille arXiv: 2609.35770

Introduction

Reconstructing realistic and editable animal fur 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.

Method Highlights

Strand-based reconstruction: FurE recovers a per-strand, editable groom by optimizing a root-conditioned latent field.
PCA decoder: The latent field is decoded into strand geometry via a PCA-based decoder.
Defurred body reconstruction: The underlying animal body is reconstructed using local fur-thickness cues from a surface-constrained Gaussian Frosting representation together with part-based priors.
Human hair transfer: A PCA-based decoder learned from human-hair strand data alleviates animal-data scarcity while significantly accelerating optimization.

Results

~10x faster strand training than current SOTA dense per-strand optimization methods.
Maintains strand fidelity while speeding up training.
Generalizes well on both synthetic and real-world sequences.