We present a system that improves the quality of noisy and incomplete depth maps captured with inexpensive range sensors. We use a model-based approach that measures the discrepancy between a model hypothesis and observed depth data. We represent the model hypothesis as a 3D level-setembedding function and the observed data as a point cloud coming from a segmented region associated to the object of interest. The discrepancy between the model and the observed data defines an objective function, that is minimized to obtain pose, scale and shape. The variation in shape of the object of interest is mapped with Gaussian Process Latent Variable Models GPLVM and the object pose is estimated using Lie algebra. The integration of a synthetic depth map, obtained from the optimal model, and the observed depth map is carriedout with variational techniques. As consequence, we work in the observed space (depth space) rather than in a high dimensional space (3D points or volumetric space), achieving outstanding results in the improved depth maps.