Evaluation of local spatio-temporal salient feature detectors for human action recognition

Local spatio-temporal salient features are used for a sparse and compact representation of video contents in many computer vision tasks such as human action recognition. To localize these features (i.e., key point detection), existing methods perform either symmetric or asymmetric multi-resolution temporal filtering and use a structural or a motion saliency criteria. In a common discriminative framework for action classification, different saliency criteria of the structured-based detectors and different temporal filters of the motion-based detectors are compared. We have two main observations. (1) The motion-based detectors localize features which are more effective than those of structured-based detectors. (2) The salient motion features detected using an asymmetric temporal filtering perform better than all other sparse salient detectors and dense sampling. Based on these two observations, we recommend the use of asymmetric motion features for effective sparse video content representation and action recognition.