Even though multiple public motor imagery classification datasets exist, normally single subject data is limited. Besides, the data acquisition process for a single individual is time-consuming, especially if the required data volume is large. This can lead to discomfort, compromising the quality of acquired data. To reduce the required data extent (and consequently its acquisition time) for training a single-subject specialized classification system, this paper aims to evaluate the application of transfer-learning in cross-subject classification of motor imagery EEG signals. A 3-class dataset, composed of 64 channel EEG acquisitions from 109 subjects, during three 2-minute sessions of motor imagery is used in the experiments. Two deep neural network architectures, widely used in previous papers, are also evaluated by comparing accuracies between models trained on single-subject and cross-subject scenarios. For both architectures, comparing single-subject and cross-subject, experiments showed an average increase of accuracy greater than 10% by using the proposed approach, resulting in an average accuracy of approximately 70%.