A method and device for denoising depth images comprises providing a noise model introducing noise to an input (13) and representing at least one three-dimensional (3D) environment with a computer graphics simulator (11). A first dataset of noiseless depth images from the represented at least one 3D environment is produced (12). The first dataset is applied as input to the noise model and the noise model producing a second dataset of noisy depth images from the first dataset (13). The first dataset and the second dataset are applied to a first machine learning system to adapt one or more filters of the first machine learning system (14).