Preface to the USTC alumni's series Preface Part Ⅰ Segmentation and Registration Chapter 1 Graph Cuts Based Active Contours (GCBAC) Chapter 2 A Novel Region Constrained Non-Rigid Image Registration FramewQrk
Part Ⅱ Face and Biometrics Chapter 3 Parallel Image Matrix Compression for Face Recognition Chapter 4 Facial Expression Recognition Based on Statistical Local Features Chapter 5 A Hierarchical Compositional Model for Face Representation and Sketching Chapter 6 A Brief Introduction to Skeleton-Based Fingerprint Minutiae Extraction
Part Ⅲ Image Annotation Chapter 7 Image Transform Bootstrapping and Its Applications to Semantic Scene Classification Chapter 8 Bipartite Graph Reinforcement Model for Web Image Annotation
Part Ⅳ Video Analysis Chapter 9 Motion Estimation Based on Trilinear and Optical Flow Constraints Chapter 10 Appearance Modeling for Visual Tracking Chapter 11 Robust Monocular 3D Tracking of Articulated Arm Movement Chapter 12 Video Classification via Local 3D Eigen Analysis Chapter 13 Video Annotation: Supervised, Semi-Supervised and Active Learning Approaches
Part Ⅴ 3D Reconstruction Chapter 14 Rapid 3D Modeling from a Single Image Based on Minimal 2D Control Points Chapter 15 Quasiconvex Optimization for Robust Geometric Reconstruction Chapter 16 Deformable Structure from Motion: A Factorization Scheme
摘要
Indeed we are far from the first who tried to patch spatial and temporal information together for a better understanding of the world: The 4D event space and light cone in special relativity put forward by Einstein have been the foundation of modern physics for about one century [34]. Even in video understanding several research groups have come to the realization of the benefits provided by patching videos in the temporal direction. Stauffer and Grimson developed the concept of pixel process , where a pixel process for a video, a 1D data, is formed by threading pixels of the same location in each frame in videos. They have applied this novel concept with valuable performance in a series of scene monitoring and tracking applications within the framework of background subtraction. Chin and colleagues proposed a temporal segmentation method based on a 2D image which is formed by stacking collinear pixels in a video. The cut detection is then transformed into a line detection problem in this 2D spatial-temporal image, which can be achieved more effectively. As can be seen these two methods entirely exploited temporal redundancies but fell short of taking full advantage of spatial redundancies. Nonetheless, the E1 time series for pixel process and E2 time frames are powerful enough to solve their respective problems elegantly. ……