Fine-grained normal estimation via feature-augmented knowledge transfer

Meng Wang

doi:10.1016/j.patcog.2026.113546

Fine-grained normal estimation via feature-augmented knowledge transfer

2026-03-19

0

PRE

AI

DOI：10.1016/j.patcog.2026.113546

Original

Original request for help

Save

Abstract

En 中文

• We introduce a geometry-aware distillation framework that transfers geometric priors from a high-capacity teacher model to a flexible student model, enabling the recovery of fine-grained surface details in normal estimation. • We propose a feature-guided consistency loss that aligns the student’s predicted normals with structural cues extracted from the input image, enhancing the preservation of high-frequency geometric information. • We show through extensive experiments that our method achieves state-of-the-art performance, particularly in retaining fine details and producing accurate, high-fidelity surface normal maps in challenging, real-world settings.

Keywords:

geometry-aware distillation

normal estimation

feature-guided consistency loss

knowledge transfer

surface normal maps

Fine-grained normal estimation via feature-augmented knowledge transfer

Abstract

Journal

Organization

Cited Papers