Sparse representation (SR) models have been studied for a long time, showing great power in acquiring, representing, compressing, restoring and even classifying high-dimensional signals. Their success is largely due to the fact that many important classes of signals including audio and images have naturally sparse representations with respect to fixed bases (i.e., Fourier, wavelet, or concatenations of them) or just some representative samples in the domain-specific subspaces. Among the enthusiasm in believing the power of sparsity in classification tasks, there is a distinctive argument which is worth noticing: “it is the collaborative representation of the test sample using all the training samples that truly results in SRC’s success but not the sparsity of representation coefficients.” This argument has motivated our recent research on collaborative representation (CR) for classification. In this area, we have got some encouraging and inspiring results, especially on the problem of person re-identification. I will introduce our progress in developing collaborative representation models for this specific application and also show that how such models perform on other classification problems.