Vlado Menkovski bio photo

Vlado Menkovski

Assistant Professor at Eindhoven University of Technology.

Email Google Scholar

Generative Adverserial Networks (GAN) have demonstrated significant success in learning the distribution of the data in an unsupervised fashion [1]. As such they have been applied in many different settings, as pre-trainer for applications with little supervised data, anomaly detection [], as well as for generating new datapoints and possibly augmenting the dataset, active learning etc. The quality of the generated data of GAN models for many types of data is typically subjectively assessed. Natural images generated by GAN models tend to be much sharper than other generative models and tend to incorporate many existing feature from the training data. However, since these models typically lack the contextual understanding of the data the generated images are still perceived with low level of ‘realness’.Subjective perception measurements are, however, difficult to incorporate in the training process of such generative models.

In this project we propose to build on top of advances in psythometics on subjective perception of stimuli [2][3] and advances in generative models in deep learning to develop a solution for generation of natural images that can benefit from feedback on subjective perception of ‘realness’ in the generated images.

[1] GAN Goodfellow paperGoodfellow, Ian, et al. “Generative adversarial nets.” Advances in neural information processing systems. 2014.

[2] Maloney, Laurence T., and Joong Nam Yang. “Maximum likelihood difference scaling.” Journal of Vision 3.8 (2003): 5-5.

[3] Menkovski, Vlado, and Antonio Liotta. “Adaptive psychometric scaling for video quality assessment.” Signal Processing: Image Communication 27.8 (2012): 788-799.Menkovski MLDS QoE paper