Stereo and ToF data fusion by learning from synthetic data.

•The paper proposes a framework for stereo and ToF data fusion.•Confidence information for the two sources is jointly estimated using a CNN.•The 2 sources are fused enforcing depth local consistency guided by confidence data.•Both synthetic and real world datasets have been created and are publicly available.•The CNN has been trained with synthetic data and is able to generalize to real data.