Resumen de Hierarchical partition-based representations of motion-coherent regions for video object segmentation
This thesis addresses the development of hierarchical partition-based representations for the extraction of video objects. Motion, as well as spatial features, from multiple frames are used to generate a set of regions structured within a hierarchy of scale and motion coherency. The resulting representations offer a global description of the entire video sequence and enhance semantic analysis potential.
A multi-scale segmentation strategy is adopted in which region merging criteria of increasing complexity are used in building partition sequences of progressively greater spatio-temporal consistency. Initially, intra-frame features are used in forming a sequence of independently generated fine color-based partitions. Following, a set of short-term features among consecutive frames are applied in the construction of a partition sequence of temporally linked, or tracked, color-homogeneous regions of coherent motion in terms of affine models. Both established metrics and metrics proposed specifically for the context of tracked partitions are applied in evaluating performance and drawing comparisons with other methods.
Long-term features, gathered across multiple frames, are key to the continued reinforcement of spatio-temporal consistency. The temporal links established among regions are used in defining concepts and tools, such as the Trajectory Adjacency Graph, for the long-term analysis of partition sequences. This analysis includes mechanisms for verifying temporal continuity and proposes merging operations with similarity measures based on global characteristics of the entire sequence. Long-term similarity measurements based on color, affine motion and translational motion are sequentially combined to produce temporally linked partition hierarchies. The transitions among different merging criteria are monitored through modeling error curves and supplemented by splitting operations.
