Recently, local descriptors have drawn a lot of attention as a representation method for action recognition. They are able to capture appearance and motion. They are robust to viewpoint and scale changes. They are easy to implement and quick to calculate. Moreover, they have shown to obtain good performance for action classification in videos. Over the last years, many different local spatio-temporal descriptors have been proposed. They are usually tested on different datasets and using different experimental methods. Moreover, experiments are done making assumptions that do not allow to fully evaluate descriptors. In this paper, we present a full evaluation of local spatio-temporal descriptors for action recognition in videos. Four widely used in state-of-the-art approaches descriptors and four video datasets were chosen. HOG, HOF, HOG-HOF and HOG3D were tested under a framework based on the bag-of-words model and Support Vector Machines.