Human action recognition using late fusion and dimensionality reduction

This paper addresses the problem of action recognition. We introduce local feature representations which are HOG, HOF, MBH, trajectory descriptor based on paper [1]. We extract those features only used one scale while Wang's paper has eight spatial scales. Our method can save memory and computation cost while guarantee the accuracy. Firstly, we apply a PCA on the HOG, HOF, MBH, trajectory descriptors to reduce the number of features. Secondly, we use Fisher kernel (FK) to aggregate each descriptor into a Fisher vector (FV) or vector of locally aggregated descriptors (VLAD) and then use improved LDA technique for FV or VLAD before being fed into the linear SVM. Thirdly, we apply late fusion for all kinds of descriptors. We evaluate our descriptor on the KTH and Youtube dataset, and as a result, observe improved performance in terms of mean average precise (mAP). Our method not only significantly reduces computational cost but improves accuracy.