Domain Adaptation for Robust Visual Recognition

Presented by Boqing Gong     

Friday, December 08, 2017
2:00 p.m.
ICSI Lecture Hall

Abstract:

During the last half decade, deep learning models have triumphed over various perception tasks in computer vision. However, to train these models requires a huge amount of data, which is difficult to collect and laborious to annotate. Some recent works thus seek alternative training cues and learn deep models from the data sources that are originally not compiled for the target task, such as Webly supervised learning, learning from photo-realistic synthetic imagery, and personalized video summarization by exploring generic videos. Despite this, the mismatch between the source and target domains significantly decreases the models’ performance.

Hence, we propose novel domain adaptation approaches to explicitly tackling the curse of domain mismatch. A key observation is that domain adaptation entails discovering and leveraging latent structures in the source and the target domains. To this end, we develop kernel methods and posterior regularization techniques for shallow models and deep neural networks, respectively. In this talk, I will give three concrete examples realizing these techniques: landmarks for effective Webly supervised learning, curriculum adaptation for deep semantic segmentation of urban scenes, and learning domain-invariant features for attribute detection. We demonstrate their effectiveness on well-benchmarked datasets and metrics.

Speaker Bio:

Boqing Gong is an Assistant Professor in Computer Science and CRCV (Center for Research in Computer Vision) at the University of Central Florida (UCF). His research lies at the intersection of machine learning and computer vision, and has been focusing on domain adaptation, zero-shot/transfer learning, and visual analytics of objects, scenes, attributes, and human activities. Boqing received his Ph.D. in Computer Science from the University of Southern California in 2015, where his work was partially supported by the Viterbi Fellowship. He holds a Master of Philosophy degree from the Chinese University of Hong Kong and a Bachelor of Engineering degree from the University of Science and Technology of China. Boqing received an NSF CRII award in 2016 and an NSF BIGDATA award in 2017, both of which were the first of their kinds granted to UCF. He actively serves on the conference program committees of computer vision (CVPR, ICCV, ECCV, etc.) and machine learning (ICML, NIPS, A! ISTATS, etc.), and was recognized as one of the Outstanding Reviewers for IEEE CVPR 2017.