On The State of Data In Computer Vision: Human Annotations Remain Indispensable for Developing Deep Learning Models