Multi-view Learning and Its Applications in Computer Vision
(Call for Papers)

      In the era of big data, multiple features and modalities are often used to describe visual data from different aspects in computer vision community. For instance, in image/video processing, different feature descriptors such as SIFT, LBP, HOG and GIST are usually adopted to represent multimedia data such as images, video frames and social media contents. In addition, various sensors can capture data from a variety of domains or modalities, such as RGB and depth data during video/image acquisition, Magnetic Resonance Imaging (MRI), positron emission tomography (PET) and Single Nucleotide Polymorphisms (SNP) in medical data processing. These data can be uniformly regarded as multi-view data. In multi-view data, different views often contain both complementary and consensus information. Fully using multi-view information is critical for enhancing learning tasks, which derives many multi-view learning methods in computer vision community. During the past few decades, although various multi-view learning methods have been put forward for computer vision tasks and gained great success, there are still many unsolved issues which need to be further investigated. For example, missing or noisy values often exist among multi-view data, and the big data era also brings the scalability problem for handling large-scale computer vision data, to name just a few.

      In addition, it is relatively easy to incorporate multiple views into deep learning, but there is much less straight forward to interpret the contributions of different views. This is unlike the shallow case (e.g., multiple kernel learning). Furthermore, the hierarchical setup of deep learning offers possibilities to learn weights at multiple levels, early and late fusion and everything in between. While some traditional multi-view learning works are well understood theoretically, there is little or no significant learning theory for deep multi-view learning. Therefore, we propose to organize this workshop which focuses on some advanced multi-view learning techniques and theory for computer vision tasks. In this workshop, we have invited several top scholars to give talks about some frontier problems of multi-view learning on computer vision.

1. Xinwang Liu, Professor, National University of Defense Technology, China, Email: xinwangliu@nudt.edu.cn;
2. Chang Tang, Professor, China University of Geosciences (Wuhan), China, Email: tangchang@cug.edu.cn;
3. Marius Kloft, Professor, TU Kaiserslautern, Germany, Email: neu@cs.uni-kl.de;
4. Xiaojun Chang, Professor, University of Technology Sydney, Australia, Email: XiaoJun.Chang@uts.edu.au;
5. Tongliang Liu, Senior Lecturer, The University of Sydney, Australia, Email: tongliang.liu@sydney.edu.au;

1. Multi-view clustering of image/video data;
2. Multi-view feature extraction of image/video data;
3. Multiple kernel learning for computer vision;
4. Multi-view classification for computer vision;
5. Multi-view representation learning for computer vision;
6. Incomplete multi-view learning for computer vision;
7. Deep multi-view learning theory;
8. Multi-view graph convolutional network for computer vision;
9. Multi-view fusion for computer vision;
10. Multi-view information retrieval for computer vision;

@ACCV 2022.