Distribution-Sensitive Learning on Relevance Vector Machine for Pose-Based Human Gesture Recognition

Many real-world gesture datasets are by nature containing unbalanced number of poses across classes. Such imbalance severely reduces bag-of-poses based classification performance. On the other hand, collecting a dataset of human gestures or actions is an expensive and time-consuming procedure. It is often impractical to reacquire the data or to modify the existing dataset using oversampling or undersampling procedures. The best way to handle such imbalance is by making the used classifier be directly aware and adapt to the real condition inside the data. Balancing class distribution, i.e., the number of pose samples per class, is one of difficult tasks in machine learning. Standard statistical learning models (e.g., SVM, HMM, CRF) are insensitive to unbalanced datasets. This paper proposes a distribution-sensitive prior on a standard statistical learning, i.e., Relevance Vector Machine (RVM), to deal with the imbalanced data problem. This prior analyzes the training dataset before learning a model. Thus, the RVM can put more weight on the samples from under-represented classes, while allows overall samples from the dataset to have a balanced impact to the learning process. Our experiment uses a publicly available gesture datasets, the Microsoft Research Cambridge-12 (MSRC-12). Experimental results show the importance of adapting to the unbalanced data and improving the recognition performance through distribution-sensitive prior.