Thursday, July 09, 2015
DeepStereo - Street View Movies
A small team at Google has published a research paper explaining how a series of still images can be processed into a seamless video. In the video above you can see how the process can be used to turn a series of Street View images into a flowing movie.
In the past we've seen projects like Street View Hyperlapse create animations from series of adjacent Street Views. While Hyperlapse is incredibly clever the results are rather like stop-motion animations. The short distance between each adjacent Street View means that the resulting animations can seem a little jarring.
Google's new DeepStereo actually creates new images to fill in the gaps between two different images, resulting in far smoother animations. As you can see in the video above the process can produce very smooth tracking shots. To 'fill in the gaps' DeepStudio analyses the depth and color from two adjacent Street View images to create a new image. The process can infer depth by comparing objects in consecutive Street View images. Nearer objects will move further from one image to the next than objects further away. DeepStudio is then able to infer the color and movement of the objects and create a new image predicting the view between the two original images.
You can imagine how Google Maps could use DeepDream to create seamless videos based on its existing Street View imagery. This would enable Google Maps, for example, to provide video previews of driving or walking routes.
However you shouldn't get too excited just yet. Each new image created by DeepDream currently takes about 12 minutes of processing. So we won't see DeepDream implemented on Google Maps anytime soon.