As an important branch of parallel vision, parallel imaging provides massive image data for parallel vision research. Specifically, a small number of real images are inputted to the software-defined artificial imaging system, where a much larger number of virtual images are generated. The small-scale real images and large-scale virtual images are combined into parallel image data, from which application-specific knowledge can be acquired through computational experiments and parallel execution. A lot of methods can be used to generate the desired virtual images, including graphics rendering, image style transfer, and generative models. In general, the generation process of virtual images is controllable, observable, and repeatable, thereby providing a useful experimental platform for studying a variety of visual computing models.
Object detection is an important topic in the computer vision field, with great theoretical and practical value in such applications as visual surveillance, autonomous driving, and human-machine interaction. In recent years, significant breakthroughs of deep learning have been attracting attention of computer vision researchers, and led to rapid development of the object detection topic. There are two main types of object detection methods: region proposal-based methods and proposal-free methods. The region proposal-based methods, especially Faster R-CNN and Mask R-CNN, have achieved high detection accuracy, whilst the proposal-free methods have superiority in detection speed. We are studying real-time methods to detect and segment objects simultaneously.
Object tracking is a research hotspot that aims at locating moving objects sequentially and obtaining their motion trajectories. This study can serve a number of purposes, including human-computer interaction, security and surveillance, video communication and compression, augmented reality, traffic control, medical imaging, and video editing. Specifically, object tracking technologies take advantages of objects’ appearance features, such as color and texture, and motion features, such as position and speed, to associate object candidates in adjacent frames. Despite its wide usages, object tracking is challenging due to severe occlusion, changing appearance, non-linear motion, and many others. We are studying robust methods to track objects in complex traffic scenes.
Computer algorithms have made great achievements in fields of image processing, voice recognition, and control engineering. Artistic creation regarded as a unique skill of human beings is becoming a hot but difficult research area. Parallel Monet is a new kind of robotic painting framework based on ACP theory (i.e., Artificial Societies, Computational Experiments, and Parallel Execution), combining the current research on image style transfer and robotic painting. The target of parallel Monet is to generate paintings of Monet’s style with the same scenes as the given real images. Different from traditional ways of end-to-end image style transfer, parallel Monet imitates the painter’s stroke and painting procedure, draws a painting stroke by stroke. The software-defined painting robot generates stroke sequence, and guides the painting procedure of physical painting robot.