A modified sports genre categorization framework based on close-up view pre-detection

In this paper a modified sports genre categorization framework is presented. The view type of close-up is detected as domain knowledge before categorization on large scale database. Close-up views occupy more than 1/3 of the duration of a sport match depending on its genre, and appears almost the same in various genres, which largely affected the performance of sports genre categorization. The presented framework is formed into two levels, a skin-tone based human detector are performed on all the key-frames to identify the close-up views in the first level. The second level is based on bag-of-word (BOW) model using Scale Invariant Feature Transform (SIFT) and Support Vector Machine (SVM) with close-ups detected in the first level. In training part, codebook is generated without close-ups according to the annotation; while in the testing part, the scores of close-ups pre-detected in the first level are calculated in low weights to make late fusion. Experiments on a dataset of 10 sports genres with 300 hours of videos from TV and Internet to ensure diversity have proven the improvements on the robustness and efficiency using our modified framework on sports genre categorization in both TV and Internet applications.