This paper presents general classes of optimal training signals for the estimation of frequency-selective channels in MIMO OFDM systems. Basic properties of the discrete Fourier transform are used to derive the optimal training signals which minimize the channel estimation mean square error. Both single and multiple OFDM training symbols are considered. Several optimal pilot tone allocations across the transmit antennas are presented and classified as frequency-division multiplexing, time-division multiplexing, code-division multiplexing in the frequency-domain, code-division multiplexing in the time-domain, and combinations thereof. All existing optimal training signals in the literature are special cases of the presented optimal training signals and our designs can be applied to pilot-only schemes as well as pilot-data-multiplexed schemes.