Machine learning has the potential to make PIV analysis faster, easier, and more robust, but the best possible way to perform this task is still an open question. Comparing various approaches to learning on the same model, multi-scale supervised learning is more accurate under ideal circumstances, but unsupervised learning may perform better in certain flow conditions, and especially in noisy conditions. Though the supervised approach may be more accurate in many cases, under sub-optimal conditions, its level of accuracy can be far more variable than that of the unsupervised model. Both approaches would seem to have their advantages. Whether one could build a model combining the accuracy of the supervised model with the robustness of the second remains to be seen.