|
|
|
| e-Pub |
Previous | 
Home
Section: New Results
Face-based Attribute Classification and Manipulation
Participants : Abhijit Das, Antitza Dantcheva, Francois Brémond.
Keywords: Face, Attribute, GAN, Biometrics
Due to the biasness of face analytic datasets, with respect to factors such as age, gender, ethnicity, pose and resolution, systems based on a skewed training dataset are bound to produce skewed results. Further, it has been exhibited in the literature [59] that such biases may have serious impacts on performance in challenging situations where the outcome is critical. In order to progress toward balanced face recognition and attribute estimation, the 1st International Workshop on Bias Estimation in Face Analytics was organized in conjunction with ECCV 2018. The workshop also organized a challenge to introduce a well-balanced dataset across multiple factors: age, gender, ethnicity, pose and resolution and requested for algorithms to estimate biases.
We proposed a Multi-Task Convolutional Neural Network (MTCNN) algorithm that jointly learned [37] gender, age and ethnicity by a loss function involving joint dynamic loss weight adjustment and was successful, as well as relatively unbiased in estimating age, gender and ethnicity. Our algorithm was found to be the best algorithm focusing the aim of the competition and the above mentioned research problem.
Generative Adversal Network (GAN)
models are autoregressive models depending on the global information, which can be potentially affected by its employment on local feature/ attribute-based erasuring. In addition, these models are typically trained depending on the maximum likelihood to find the intense difference between the regression domains, as a result after a certain limit of learning it can produce very naive development in the interpolation of the regression carried out for the purpose of local attribute removal. Hence, to mitigate an aforementioned couple of pitfalls we propose a method for localizing the Cycle GAN (C-GAN) for local feature-based regression. We trained the C-GAN with domain-specific local feature and end model was recurrently imposed on the testing images. We experimented the Local C-GAN (L-C-GAN) on facial attribute (eyeglass and moustache/ bearded) auto-regression. Our qualitative performance on partial CelebA dataset and a couple of datasets we collected is promising. Moreover, ensuring the facial attributes have also been found to achieve better performance accuracy with respect to the presence of these attributes.