Compensating inaccurate annotations to train 3D facial landmark localization models

Abstract

Abstract-In this paper we investigate the impact of inconsistency in manual annotations when they are used to train automatic models for 3D facial landmark localization. We start by showing that it is possible to objectively measure the consistency of annotations in a database, provided that it contains replicates (i.e. repeated scans from the same person). Applying such measure to the widely used FRGC database we find that manual annotations currently available are suboptimal and can strongly impair the accuracy of automatic models learnt therefrom. To address this issue, we present a simple algorithm to automatically correct a set of annotations and show that it can help to significantly improve the accuracy of the models in terms of landmark localization errors. This improvement is observed even when errors are measured with respect to the original (not corrected) annotations. However, we also show that if errors are computed against an alternative set of manual annotations with higher consistency, the accuracy of the models constructed using the corrections from the presented algorithm tends to converge to the one achieved by building the models on the alternative, more consistent set.

3-D Facial Landmark Localization With Asymmetry Patterns and Shape Regression from Incomplete Local Features

“…The ground truth displacements are only used during training to derive the templates and are specific to each descriptor. We have shown that this strategy is more accurate than simply trusting the manual annotations [57].…”

“…In all cases, we obtained descriptor templates for each landmark by averaging over the training set. As manual annotations for FRGC have been shown to be rather noisy, we used the least squared corrections of uncertainty algorithm [57] to build the templates. In brief, this means that we assumed an uncertainty in the manual annotations, which were allowed to move within a small neighborhood of radius r u to enforce consistency of the extracted descriptors.…”

“…The experiments presented in the previous sections aimed at testing the robustness of SRILF and its performance in relation to state of the art approaches using large public databases in which there are acquisition artifacts, occlusions, strong facial expressions, and a nonnegligible degree of noise in the manual annotations, as discussed in [57]. In this section, we explore how much can we reduce localization errors by testing SRILF on a clinical dataset where special care has been taken to minimize the presence of artifacts and manual annotations have been performed by experts.…”