Dynamic Scene Recognition

Many applications need to deal with a scene that evolves over time. This thesis addresses the dynamic scene recognition problem. Video-based dynamic scene recognition methods and mid-level representations are comprehensively reviewed, followed by several proposed methods. First, a part-based approach for dynamic scene recognition is presented. In this approach, representative parts are located by clustering local convolutional features extracted using a pre-trained Fast RCNN model. A set cover problem is then formulated to select the discriminative parts, which are subsequently refined. Local features from a video segment can be extracted and aggregated. Extensive experiments show that this approach is very competitive with state-of-the-art methods that use global features, achieving an overall accuracy of 91.5% and 98.3%, respectively, on two benchmark datasets— the Maryland and the Yupenn.