Failure Detection for Facial Landmark Detectors

TL;DR

This paper develops failure detection methods for facial landmark detectors, improving reliability in face analysis applications by identifying inaccuracies and reducing errors in downstream tasks.

Contribution

It introduces confidence models for recent landmark detectors and demonstrates effective failure detection and error reduction on benchmark datasets.

Findings

Detects over 40% of landmark detection failures

Achieves 12% error reduction in gender estimation

Validates methods on AFLW and HELEN datasets

Abstract

Most face applications depend heavily on the accuracy of the face and facial landmarks detectors employed. Prediction of attributes such as gender, age, and identity usually completely fail when the faces are badly aligned due to inaccurate facial landmark detection. Despite the impressive recent advances in face and facial landmark detection, little study is on the recovery from and detection of failures or inaccurate predictions. In this work we study two top recent facial landmark detectors and devise confidence models for their outputs. We validate our failure detection approaches on standard benchmarks (AFLW, HELEN) and correctly identify more than 40% of the failures in the outputs of the landmark detectors. Moreover, with our failure detection we can achieve a 12% error reduction on a gender estimation application at the cost of a small increase in computation.