Face Attribute Classification & Attribute-Based Classifiers

 

We have been working on attribute-based classifiers that allow describing or searching an object by its attributes. We are working in particular on face and plant attributes and zero-shot learning.

 

 

Recent Publications:

 

§  Sara Atito Ali Ahmed, Berrin A. Yanikoglu:
Within-Network Ensemble for Face Attributes Classification. ICIAP (1) 2019: 466-476

 

Building an ensemble of deep learners is very time consuming and the benefits often not very significant since the ensembles are often highly correlated. In this work we propose a within-network ensemble, that is an ensemble of 5 networks trained within one deep network, by adding extra output layers to 4 earlier layers.

 

 

§  Sara Atito Aly, Berrin A. Yanikoglu:
Multi-Label Networks for Face Attributes Classification. ICME Workshops 2018: 1-6

 

In this work, we explore the benefits of ‘pointing to an attribute’ (indicating rough location) while training the attribute classifier. For instance, in training for smiley/not smiley, we tell the network that it involves the mouth region by blurring the image except for the mouth region. Results show that pointing speeds up learning, but requires the training to continue with the original image for max performance.