Automatic Generation of 3D GIFs

We combine monocular depth estimation with motion segmentation techniques to produce split-depth GIFs (GIF is a common animated image format, and split-depth is a technique to enhance the 3-D effect of the GIF by inserting white bars that split the depth of the scene). We use a deep convolutional neural field (DCNF) model to perform monocular depth estimation, and also perform motion segmentation using Gaussian Mixture Model for color cues and Optical Flow for motion cues into a Markov Random Field. We test the DCNF model on the NYU v2 dataset and find an average relative error of 0.24 and a RMS error of 0.825. We also compare the produced split-depth GIFs with a dataset of manually-annotated ones to find an average recall of 0.94 and average precision of 0.83 in computing the foreground/background pixels relevant to the split-depth level. We find that combining motion segmentation with monocular depth estimation results in an increase in precision of roughly 20% compared to solely using monocular depth estimation to generate the split-depth GIFs.