The Menpo Facial Landmark Localisation Challenge: A Step Towards the Solution

Abstract

In this paper, we present a new benchmark (Menpo benchmark) for facial landmark localisation and summarise the results of the recent competition, so-called Menpo Challenge, run in conjunction to CVPR 2017. The Menpo benchmark, contrary to the previous benchmarks such as 300-W and 300-VW, contains facial images both in (nearly) frontal, as well as in profile pose (annotated with a different markup of facial landmarks). Furthermore, we increase considerably the number of annotated images so that deep learning algorithms can be robustly applied to the problem. The results of the Menpo challenge demonstrate that recent deep learning architectures when trained with the abundance of data lead to excellent results. Finally, we discuss directions for future benchmarks in the topic.

The 3D Menpo Facial Landmark Tracking Challenge

“…Currently, it is feasible to robustly train DCNNs for face alignment, since our group has provided large scale landmark annotations [17,16,18,24]. In the first challenge, i.e.…”

“…The 300W and 300VW benchmarks provided annotations with regards to a frontal face shape of 68 landmarks. A step forward was made in CVPR 2017 by our group in the so-called Menpo Challenge [24]. In Menpo challenge we provided annotations for over 12,000 faces including, for the first time, annotations for over 4,000 profile faces (with regards to to 39 landmarks).…”

Learning to find eye region landmarks for remote gaze estimation in unconstrained settings

“…The stacked multiscale architecture is simple and has been shown to out-perform other state-of-the-art methods while having low model complexity (few number of parameters). Originally developed for pose estimation, the architecture has been successfully adapted to the task of facial landmark localization in the new Menpo Facial Landmark Localisation Challenge [Yang et al 2017;Zafeiriou et al 2017].…”

A Fully Automated System for Sizing Nasal PAP Masks Using Facial Photographs

“…One final dataset, described in more detail below was artificially generated using both the MUCT and PUT datasets in addition to photographs of both the head and tails side of an Australian 20 cent piece, downloaded from the internet. Samples from the MUCT and PUT datasets were selected for use during this study, as opposed to others databases such as 300W [13] and menpo [14] due to their similarity to the data collected during this study. The MUCT dataset, while containing images of approximately the same pose, with different lighting is known for being composed of participants from a variety of demographics; with both sexes, a range of different ages and ethnicities being represented in the set.…”