Abstract-The recognition of human actions from videos has several interesting and important applications, and a vast amount of different approaches has been proposed for this task in different settings. Such approaches can be broadly categorized in model-based and model-free. Typically, model-based approaches work only in very constrained settings, and because of that, a number of model-free approaches appeared in the last years. Among them, those based in bag-of-visual-features (BoVF) have been proving to be the most consistently successful, being used by several independent authors.For videos to be represented by BoVFs, though, an important issue that arises is how to represent dynamic information. Most existing proposals consider the video as a spatio-temporal volume and then describe "volumetric patches" around 3D interest points. In this work, we propose to build a BoVF representation for videos by collecting 2D interest points directly. The basic idea is to gather such points not only from the traditional frames (xy planes), but also from those planes along the time axis, which we call the spatio-temporal frames. Our assumption is that such features are able to capture dynamic information from the videos, and are therefore well-suited to recognize human actions from them, without the need of 3D extensions for the descriptors. In our experiments, this approach achieved stateof-the-art recognition rates on a well-known human actions database, even when compared to more sophisticated schemes.