eduzhai > Applied Sciences > Engineering >

Disentanglement for Discriminative Visual Recognition

  • Save

... pages left unread,continue reading

Document pages: 37 pages

Abstract: Recent successes of deep learning-based recognition rely on maintaining thecontent related to the main-task label. However, how to explicitly dispel thenoisy signals for better generalization in a controllable manner remains anopen issue. For instance, various factors such as identity-specific attributes,pose, illumination and expression affect the appearance of face images.Disentangling the identity-specific factors is potentially beneficial forfacial expression recognition (FER). This chapter systematically summarize thedetrimental factors as task-relevant irrelevant semantic variations andunspecified latent variation. In this chapter, these problems are casted aseither a deep metric learning problem or an adversarial minimax game in thelatent space. For the former choice, a generalized adaptive (N+M)-tupletclusters loss function together with the identity-aware hard-negative miningand online positive mining scheme can be used for identity-invariant FER. Thebetter FER performance can be achieved by combining the deep metric loss andsoftmax loss in a unified two fully connected layer branches framework viajoint optimization. For the latter solution, it is possible to equipping anend-to-end conditional adversarial network with the ability to decompose aninput sample into three complementary parts. The discriminative representationinherits the desired invariance property guided by prior knowledge of the task,which is marginal independent to the task-relevant irrelevant semantic andlatent variations. The framework achieves top performance on a serial of tasks,including lighting, makeup, disguise-tolerant face recognition and facialattributes recognition. This chapter systematically summarize the popular andpractical solution for disentanglement to achieve more discriminative visualrecognition.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×