Abstract-The lack of accurate and efficient channel estimation (CE) for the multi-input multi-output (MIMO) channel state information (CSI) has long been the stumbling-block of near-MIMO-capacity operation. We propose a semi-blind joint CE and three-stage iterative detection/decoding scheme for near-capacity MIMO systems. The main novelty is that our decision-directed (DD) CE exploits the a posteriori information produced by the MIMO soft-demapper within the inner turbo loop for selecting a 'just-sufficient-number' of high-quality detected soft bit blocks or symbols for DD channel estimation (DDCE), which significantly improves the accuracy and efficiency of DDCE. Moreover, our DDCE is naturally embedded into the iterative three-stage detection/decoding process, without imposing an additional external iterative loop between the DDCE and the three-stage turbo detector/decoder. Hence, the computational complexity of our joint CE and three-stage turbo detector/decoder remains similar to that of the three-stage turbo detection/decoding scheme associated with the perfect CSI. Most significantly, the mean square error of our DD channel estimator approaches the Cramér-Rao lower bound associated with the optimal training based CE, while the bit error rate of our semi-blind scheme is capable of achieving the optimal maximum-likelihood performance bound associated with the perfect CSI.Index Terms-Multi-input multi-output, near-capacity, joint channel estimation and three-stage turbo detection/decoding, Cramér-Rao lower bound