Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PaW-ViT: A Patch-based Warping Vision Transformer for Robust Ear Verification

Deeksha Arun, Kevin W. Bowyer, Patrick Flynn

TL;DR

This work introduces PaW-ViT, an anatomy-guided preprocessing pipeline for ear verification that transforms ear geometry into a structured, anatomy-preserving 112×112 canvas by sampling boundary points, constructing a triangular fan, forming quadrilaterals, and affine-warping into 16 fixed patches. By aligning patch boundaries to natural ear curves and stitching patches into a grid, PaW-ViT preserves both local and global ear features while suppressing background noise, enabling ViTs to learn more stable embeddings. Evaluations across four datasets (OPIB, AWE, WPUT, EarVN1.0) and four ViT configurations (ViT-T, ViT-S, ViT-B, ViT-L) show larger models generally perform best, with union and intersection map-based warping yielding the strongest gains on challenging data—especially EarVN1.0—while segmentation- and landmark-based warps alone offer limited improvements. The approach highlights the potential of geometry-driven preprocessing to bridge the gap between morphological variability in biometrics and transformer positional sensitivity, offering a practical path toward robust, anatomy-aware ear recognition.

Abstract

The rectangular tokens common to vision transformer methods for visual recognition can strongly affect performance of these methods due to incorporation of information outside the objects to be recognized. This paper introduces PaW-ViT, Patch-based Warping Vision Transformer, a preprocessing approach rooted in anatomical knowledge that normalizes ear images to enhance the efficacy of ViT. By accurately aligning token boundaries to detected ear feature boundaries, PaW-ViT obtains greater robustness to shape, size, and pose variation. By aligning feature boundaries to natural ear curvature, it produces more consistent token representations for various morphologies. Experiments confirm the effectiveness of PaW-ViT on various ViT models (ViT-T, ViT-S, ViT-B, ViT-L) and yield reasonable alignment robustness to variation in shape, size, and pose. Our work aims to solve the disconnect between ear biometric morphological variation and transformer architecture positional sensitivity, presenting a possible avenue for authentication schemes.

PaW-ViT: A Patch-based Warping Vision Transformer for Robust Ear Verification

TL;DR

This work introduces PaW-ViT, an anatomy-guided preprocessing pipeline for ear verification that transforms ear geometry into a structured, anatomy-preserving 112×112 canvas by sampling boundary points, constructing a triangular fan, forming quadrilaterals, and affine-warping into 16 fixed patches. By aligning patch boundaries to natural ear curves and stitching patches into a grid, PaW-ViT preserves both local and global ear features while suppressing background noise, enabling ViTs to learn more stable embeddings. Evaluations across four datasets (OPIB, AWE, WPUT, EarVN1.0) and four ViT configurations (ViT-T, ViT-S, ViT-B, ViT-L) show larger models generally perform best, with union and intersection map-based warping yielding the strongest gains on challenging data—especially EarVN1.0—while segmentation- and landmark-based warps alone offer limited improvements. The approach highlights the potential of geometry-driven preprocessing to bridge the gap between morphological variability in biometrics and transformer positional sensitivity, offering a practical path toward robust, anatomy-aware ear recognition.

Abstract

The rectangular tokens common to vision transformer methods for visual recognition can strongly affect performance of these methods due to incorporation of information outside the objects to be recognized. This paper introduces PaW-ViT, Patch-based Warping Vision Transformer, a preprocessing approach rooted in anatomical knowledge that normalizes ear images to enhance the efficacy of ViT. By accurately aligning token boundaries to detected ear feature boundaries, PaW-ViT obtains greater robustness to shape, size, and pose variation. By aligning feature boundaries to natural ear curvature, it produces more consistent token representations for various morphologies. Experiments confirm the effectiveness of PaW-ViT on various ViT models (ViT-T, ViT-S, ViT-B, ViT-L) and yield reasonable alignment robustness to variation in shape, size, and pose. Our work aims to solve the disconnect between ear biometric morphological variation and transformer architecture positional sensitivity, presenting a possible avenue for authentication schemes.
Paper Structure (18 sections, 4 figures, 1 table)

This paper contains 18 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Ear Image Preprocessing
  5. Ear Boundary Extraction
  6. Ear Boundary Sampling and Centroid Computation
  7. Triangular Fan Construction
  8. Quadrilateral Formation and Affine Warping
  9. Patch Stitching
  10. Datasets
  11. Training data
  12. Testing data
  13. Experimental Setup
  14. Results
  15. Model-based Performance Results
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Landmark map generation process: (a) input ear image, (b) detected landmark points using 2SHG-Net model hrovativc2023efficient, (c) landmark map overlay on the ear image generated from the detected landmark points, and (d) final landmark map.
  • Figure 2: Maps used in the patch-based warping process: (a) original ear image, (b) segmentation map, (c) landmark map, (d) intersection map, and (e) union map obtained using the landmark and segmentation maps.
  • Figure 3: Detailed methodology of the PaW-ViT algorithm. (Note that the green lines and text-labels are added only to illustrate the flow between the steps, and they are not part of the method.)
  • Figure 4: Peak performance of ViT models across datasets. Each bar's label indicates the maximum AUC score and the specific map type that achieved it. The AUC score (and its label) for the top-performing model for each dataset is highlighted in red.