Single image depth estimation using multi-scale deep connections and batch data augmentation

Depth estimation is traditionally done using stereo images. Recently depth has been estimated from a single image using convolutional neural networks at different scales. This work builds on these developments with two enhancements. First we show that adding deep connections across the three scales in a multi-scale setup improves the depth estimate. Second, we show that augmenting each batch data with both original and horizontally flipped images and passing them through the same layers, helps to further improve the depth estimate. Experimental results on the NYUD dataset validate these enhancements.