Learning effective RGB-D representations for scene recognition