Special Issue on Compositional Models and Structured Learning for Visual Recognition

  In the last ten years, computer vision and pattern recognition has experienced a resurgence of research on compositional and hierarchical models, such as And-Or graphs, deformable part-based models, kernelized and latent variable models. The virtue of compositional and hierarchical models (CHMs) lies in their expressive power to model diverse and complex visual patterns. Meanwhile, a set of structured learning and optimization methods are intensively discussed to facilitate training and inference with compositional models, which usually integrate latent structures to specify the task-specific compositional configurations and contextual relations. These methods, such as latent support vector machines, conditional random fields, and structural sparse coding, enable inference with rich internal structures and pursue a good mapping between observations and output structured predictions. Compared with the neural networks, which have also attracted much attention recently, CHMs and structured learning methods provide alternative approaches to explicitly handle the variations of data with latent variables, and demonstrate their potential in several high-level vision tasks, e.g., object detection and recognition, scene parsing, and action/activity understanding.

In order to pursue first-class research along this direction, we would like to organize a special issue titled "Compositional Model and Structured Learning for Visual Recognition" in the journal of Pattern Recognition.

The issue will be aimed at accepting papers on the following topics but not limited to:

1. Object recognition and detection by learning and inference with compositional and hierarchical models. The proposed approaches are encouraged to evaluate on several public benchmarks in computer vision such as PASCAL VOC, ImageNet, Caltech101, and Caltech256.
2. Image segmentation and labeling with supervised or unsupervised learning methods, which incorporate multiple contextual object models. Some exemplar benchmarks are LabelMe, PASCAL VOC, Fashionista?, and SUN databases.
3. Understanding human actions or activities from videos with spatio-temporal models. The new models will show good performance on capturing well large category variations that is one key challenge in complex action/activity modeling. By using depth sensors, more rich information can be utilized for these tasks.
4. Models, algorithms, and applications of sparse representation and dictionary learning. The proposed approaches are expected to improve the efficiency and effectiveness of the classification performance, and provide new insight for modeling structure and dependencies between vocabularies.
5. New applications and systems address real challenges in the intelligent processing and understanding of visual data (e.g. fashion understanding, medical image recognition, graphics, etc).?

The main timelines for this issue are set as follows,

• Paper submission due: July. 30, 2015
• First notification: Nov. 30, 2015
• Revision: Jan. 15, 2016
• Final decision: Feb. 30, 2016

Submission Details:

All submissions for this special issue are required to follow the same format as regular full-length Pattern Recognition papers. The submission website for this special issue is located at: http://ees./pr/. Please ensure to select 'SI : CHM-Vision' as the 'Article Type'.

Guest Editors:

Professor Liang Lin
Sun Yat-Sen University
[e-mail]

Associate Professor Jason Corso
University of Michigan
[e-mail]

Associate Professor Wangmeng Zuo
Harbin Institute of Technology
[e-mail]

Chair Professor David Zhang
The Hong Kong Polytechnic University
[e-mail]

Dr Benjamin Yao
Amazon.com
[e-mail]