Methods based on feature descriptors around local interest points are now widely used in action recognition. Feature points are detected using a number of measures, namely saliency, periodicity, motion activity etc. Each of these measures is usually intensity-based and provides a trade-off between density and informativeness. In this paper, we address the problem of action recognition by representing image sequences as a sparse collection of patch-level space-time events that are salient in both space and time domain. Our method uses a multi-scale volumetric representation of video and adaptively selects an optimal space-time scale under which the saliency of a patch is most significant. The input image sequences are first partitioned into non-overlapping patches. Then, each patch is represented by a vector of coefficients that can linearly reconstruct the patch from a learned dictionary of basis patches. The space-time saliency of patches is measured by Shannon’s self-information entropy, where a patch’s saliency is determined by information variation in the contents of the patch’s spatiotemporal neighborhood. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed method.