A short Two Minute Papers video (2016) explaining research that estimates depth maps from a single photograph. The presenter describes collecting large datasets of image–depth pairs (e.g., by mounting 3D scanners on a vehicle), shows side-by-side examples of input images, ground-truth depth heatmaps, and predicted depth heatmaps, and highlights applications such as novel-view synthesis and low-cost robot navigation.
Chapters
Dear fellow scholars, this is two minute papers. With this piece of work tries to estimate depth information from an input photograph. This means that it looks at the photo and tries to tell how far away parts of the image are from the camera. An example output looks like this. On the left there is an input photograph. And on the right you see a heatmap with true distance information. This is what we are trying to approximate. This means that we collect a lot of indoor and outdoor images with their true depth information. And we try to learn the correspondence, how they relate to each other, sidewalks, forests, buildings, you name it. These image and depth pairs can be Captured by mounting 3D scanners on this awesome custom built vehicle. And gentlemen, that is one heck of a way of spending research funds. The final goal is that we provide a photograph for which the depth information is completely unknown. And we ask the algorithm to provide it for us. Here you can see some results. The first image is the input photograph. The second shows the true depth information. The third image is the depth information that was created by this technique. And here's a bunch of results. For images downloaded from the Internet, it probably does at least as good as a human would. Spectacular. This sounds like a sensorial problem for humans and a perilous journey for computers to say the least. What is quite remarkable is that these relations can be learned by a computer algorithm. What can we use this for? Well, a number of different things. One of which is to create multiple views of this 2D photograph using the guessed depth information. It can also be super helpful in building robots that can wander about reliably with inexpensive consumer cameras mounted on them. Thanks for watching and for your generous support and I'll see you next time.
