Exploiting Web Images for Semantic Video Indexing Via Robust Sample-Specific Loss

摘要: Semantic video indexing, also known as annotation or concept detection in literatures, has been attracting significant attention recent years. Due to deficiency of labeled training videos, most the existing approaches can hardly achieve satisfactory performance. In this paper, we propose a novel semantic indexing approach, which exploits abundant user-tagged Web images help learn robust classifiers. The following two major challenges are well studied: 1) noisy with imprecise and/or incomplete tags; and 2) domain difference between videos. Specifically, first apply non-parametric approach estimate probabilities being correctly tagged confidence scores. We then develop transfer (RTVI) model reliable classifiers from limited number videos together abundance images. RTVI is equipped sample-specific loss function, employs score image prior knowledge suppress influence control contribution learning process. Meanwhile, discovers an optimal kernel space, mismatch minimized for tackling problem. Besides, devise iterative algorithm effectively optimize proposed theoretical analysis on convergence provided well. Extensive experiments various real-world multimedia collections demonstrate effectiveness approach.