Preprocessing of 2.5D Facial Images from Stereo Camera

Authors: Myrah Naeem

Affiliation: Sejong University / Kyung Hee University

Short Overview

This work focuses on improving the quality of 2.5D facial images captured from a stereo camera for downstream face recognition tasks. The preprocessing pipeline includes median filtering, Laplacian smoothing, hole filling, cropping, and geometric alignment of point clouds. These operations remove spikes, reduce noise, smooth surface irregularities, fill missing data, and produce consistent 2.5D facial mesh maps suitable for curvature-based 3D facial feature extraction.

📄 Download Full Paper

Abstract

This paper presents a preprocessing pipeline for 2.5D facial images acquired using a stereo camera. The goal is to enhance mesh quality and prepare surface data for 3D facial recognition. The pipeline applies median filtering to eliminate noise, Laplacian smoothing to remove spikes and surface irregularities, and hole filling to address missing depth information. Additionally, cropping and facial alignment techniques are applied to isolate the face region and normalize its orientation. The resulting 2.5D facial meshes exhibit improved surface consistency and structural accuracy, enabling reliable curvature-based analysis in 3D face recognition applications.