In this paper, we propose a two-stage method to estimate 3D human pose, which focuses on the uncertainty of lifting 2D detected pose to 3D pose. Firstly, a novel category map is introduced to predict the ordinal depth category which depicts three kinds of depth ordering relationship for linked joints. Compared with the common probability of vector, our category map can provide better association between prediction with image appearance, and lead to a higher classification accuracy. Secondly, taking predicted 2D pose and ordinal depth category as input, we put forward a temporal convolution network to regress 3D pose, which exploits the temporal context to alleviate the 2D-to-3D uncertainty and reduce prediction errors rate from single image further. Experimental results show that our method can outperform promising results on several benchmarks.