Wednesday, July 3, 2013

CVPR brain dump, part 3 of n: graphs of photos

Someday, there's going to be a world-wide geo-located network of photos called "The Photo Graph". Adam and I are going to make it. Every node will be a photo with a location and orientation on the globe, and a confidence for that pose/orientation. Photos that see something in common will be connected by edges that encode matching visual features between photos and the relative change in pose.
This is a graph of facebook friends, but imagine geolocated photos and their "friends"...

The important features of the Photo Graph will be:

  • It's easy to add new photos to the photo graph, whether they go into a heavily populated area or mildly populated area or a totally new, unpopulated area
  • It's easy to update the graph in the presence of new information
  • It's easy to correct mistakes in the graph (bad edges, edges that should be there but aren't)
  • It's easy to query the graph to get photos (and matches) to use for other purposes, like on-the-fly 3D reconstruction or geolocalization 
  • It's easy to attach other useful metadata to nodes and edges in the graph, such as tags about the photographer, or the content of the photo, or human-defined "architectural" points that match between photos that a computer would not have realized were matched.
Actually a graph of photos. This shows some PhotoCity photos and matches between photos. There were more matches than Fusion Tables could cope with, though. Also this graph is totally static and can't be grown/pruned/fixed/changed in any way.

How does this pertain to CVPR? 


I didn't even talk about the photo graph at CVPR, I just stalked other people's graphs of photos.

Here are some notable graphs that were spotted:

Not a graph, just geolocalized pictures on flickr from Crandall's Mapping the World's Photos

Also from Crandall's Mapping the World's Photos, but this time the graph just shows the visualization of a photographer's movement around Manhattan

A graph of photos of Dubrovnik from Cao & Snavely's Graph-Based Discriminative Learning for Location Recognition. The edges correspond to images sharing visual features. The layout of the graph doesn't reflect how the images' spatial location in the world.

The Angled Graph thing, which doesn't encode spatial information but does make a nice representation for a particular "CrowdCam" application. Video here!

Marc Pollefeys was giving a talk about many projects, including one on Discovering and Exploiting 3D Symmetries in Structure from Motion and he had a figure of a graph of photos with some broken and incorrect links. Since I can't find a relevant image, I'm just going to use my own. This is my own figure that shows the matched photos around a particular model. You can see on the left that there's a gap in the photo matches; they really should connect and be solid pink all the way around. This symmetry exploitation approach could really help this model.

Graphs graphs graphs! So obviously graphs of photos aren't new, but I think the Photo Graph could be something new and special that borrows ideas from these other graphs and extends the things you can do to a graph and with a graph.

1 comment:

  1. Such as almost any artisan, professional photographers should hold their particular talent well-defined. More than decade as
    φωτογράφος γάμων a expert photographer, We've needed to locate solutions to maintain your inventive mindset going.

    ReplyDelete