Self-Supervised Online Metric Learning With Low Rank Constraint for Scene Categorization

摘要: Conventional visual recognition systems usually train an image classifier in a bath mode with all training data provided advance. However, many practical applications, only small amount of samples are available the beginning and more would come sequentially during online recognition. Because characteristics could change over time, it is important for to adapt new incrementally. In this paper, we present metric learning method address scene problem via adaptive similarity measurement. Given number labeled followed by sequential input unseen testing samples, learned maximize margin distance among different classes samples. By considering low rank constraint, our model not can provide competitive performance compared state-of-the-art methods, but also guarantees convergence. A bi-linear graph defined pair-wise similarity, sample depending on graph-based label propagation, while self-update using confident With ability learning, methodology well handle large-scale streaming video incremental self-updating. We evaluate categorization experiments various benchmark datasets comparisons methods demonstrate effectiveness efficiency algorithm.