Accurate electroencephalography (EEG) classification of finger flexion imagery would endow non-invasive brainmachine interfaces (BMIs) with a much richer control repertoire. Traditionally, it has been assumed that non-invasive methods lack the resolution required for success on such a fine discrimination task. In this study, we challenged this assumption. EEG was acquired while subjects imagined performing individual and bimanual finger flexions. A new method of spatiotemporal and spectral feature extraction was applied, and multi-class support vector machine (SVM) classifiers were trained. Predictions and probabilities then served as inputs to a novel voting scheme, which output the system decision. The present approach achieved a mean population (n=15) accuracy of 30.86±1.76%, nearly twice the chance guessing level (16.71±1.68%) for the six-class task evaluated. Finger imagery is thus shown to be classifiable through EEG analysis alone.