0

I'm wondering what is the best mechanism to take images from the smartphone camera in order to match theses images with my data-set ,in addition I'm planing to use SURF algorithm for features extraction & descriptor .when i should take the training set how i should take pictures (i.e in height or width manner or both ).also how many pictures in average may i need for each interest building ? I will be very thankful if someone could guide me .

Abo gaseem
  • 5
  • 5
  • i dont know why it's not clear ,I'm asking how many pictures i need in average for each building and how to capture the pictures ?in width or height or both direction – Abo gaseem Nov 10 '14 at 18:36
  • Well, a building can be a small house, a sky high skyscraper, a massive shopping mall, detail might vary from basic shape of the building to decals on the wall or stucco ornaments, do you really expect an "average number"? – dtech Nov 10 '14 at 18:39
  • yup u are true today I take some there's some cannot be taken due to the scale , I'm new to CV . in your opinion what is the best strategy i can follow ? – Abo gaseem Nov 10 '14 at 18:46

1 Answers1

1

How many pictures in average may i need for each interest building?

The short answer is "enough". Buildings vary in size, shape, detail and whatnot. You should implement some sort of algorithm which detects and registers features and should feed it new data until you don't detect any new features and only old and already detected features, you may have a "learned" factor in percent, starting from 0% and all the way to 100% when N new images no longer detect any new feature. Then you may tune that to say 70% or whatever, depending on how much time and resources you want to dedicate to creating a feature profile for a given building.

Then for some building it may take 5 images, for another building it may take 500 images. But you will have an algorithm which knows when it is enough.

dtech
  • 47,916
  • 17
  • 112
  • 190
  • @Abogaseem: In addition you might also be interested in the work made by the Mobile Visual Search team at Stanford: see their [dataset](http://msw3.stanford.edu/~dchen/datasets.html#mvsdataset) that includes landmarks, but also their [detailed paper](http://msw3.stanford.edu/~dchen/datasets.html#sflandmarkdataset) that explains how they acquired 1.7 million images of buildings in San Francisco. – deltheil Nov 11 '14 at 19:13