Ankita Somani1, Bharati Sonawane2, Amruta Shingare3, Nikita Gaikwad4

1(Department Of Computer Engineering, VPCOE, Baramati)

2(Department Of Computer Engineering, VPCOE, Baramati)

Introduction

Web videos present numerous variations in faces due to hairstyle, skin tone, wrinkles, and more, making face labeling a challenging task. Most search engines index videos with user-provided text descriptions, which are often noisy and incomplete, leading to unsatisfactory search performance. This paper introduces an algorithm based on Conditional Random Field (CRF) that utilizes relationships rather than rich text for face naming in web videos.

By leveraging visual similarities and spatiotemporal contexts, this approach addresses the issues of incomplete metadata and noisy labels, providing a solution for unsupervised face labeling in videos without prior domain-specific knowledge.

Existing System

Existing face naming research primarily focuses on Web images and constrained videos such as TV serials, news, and movies. Approaches fall into three categories:

• Model-based: Uses classifiers for face recognition but struggles to scale due to the need for labeled samples for each face model.
• Instance-based: Investigates face-name associations with binary labels (correct or incorrect) but lacks complete data knowledge.
• Unsupervised learning: Employs weakly labeled celebrity images and social networks, encoded using CRF for improved label inference.

Abstract

This paper addresses celebrity face naming in unconstrained videos using client-provided metadata. Instead of supervised learning with exact face names, the algorithm relies on relationships derived from video content, image areas, and social cues for unsupervised labeling. Two face annotation versions are explored: within-video and between-video labeling. The approach rectifies metadata errors, corrects false names, and clarifies missing names by considering socially associated videos for joint inference.

Keywords:

Face naming, social network, Web videos, within-video face naming, between-video face naming