In this paper, a Bayesian mixture expert (BME) framework for the estimation of 3D human poses from two uncalibrated wide-baseline cameras is presented. The two cameras will reduce the ambiguities of the pose estimation greatly and is easy to implemented. BME is learnt to conductmulti-modal pose estimation regression. K-means algorithm considering Euclidean distance and maximum-value distance for the joint angle vector is used for the initial clustering in BME learning. This will give the better cluster results to separate the ambiguous poses into different expert. Also a weighted PCA is implemented in an expectation-maximization (EM) framework to learn the parameters of the BME. This can reduce the dimension of the training data more effectively compared with global PCA. The system is trained with synthesized silhouettes from motion capture data. The experimental results on synthesized and real images illustrate that our approach does not need precise camera calibration and can estimate the poses effectively.