[ WEEK3– Predicting the location of o photograph ]

This week, we collected remaining part of our dataset and made investigation about this project and computer vision. Firstly, looking for what features we can use was educational for us. We will describe and give information about them.

Features can be represented with descriptors. For example one descriptors can represent features with one vector or another one can represent with more than one vector. There are two method to represent features: global features and local features. Global features are consists of properties of whole image which can be color, texture, shape etc.. However, local features are related with local areas. The main aim of local features is finding salient regions and use them against viewpoint and illumination changes.

Advantage of global features:

  • faster and compact
  • easy to compute
  • small require of memory

Disadvantage of global features:

  • Not sensitive compare to local features

Advantage of local features:

  • Perfect performance for large-scale image search
  • More convenient for image matching

Disadvantage of local features:

  • More require of memory than global features

Here, some descriptors :

  • HOG (Histogram of Oriented Gradients ) : It useus local gradients of an image to create a feature vector. First, algorithm divedes image to small cells and computes gradients. Then it distinguises cells using gradients. After normalization the vector becomes feature vector
  • SIFT (Scale Invarient Feature Transform) : This descriptor uses local gradients and histograms. Firstly, the descriptor takes image gradients and then create 128-length vector wtih histograms. It is generally used for orientation, scale problems and smooth changes. It can be useful for our project because images of city places can be taken in several viewpoints and scales. It is good solution for these problems but computational time can be very huge.
  • SURF (Speeded-Up Robust Features Descriptor) : It is an alternative of SIFT descriptor.

We are planning to construct our architecture next week. We will use deep learning as we said and our design will be shared in next blog.

One clap, two clap, three clap, forty?

By clapping more or less, you can signal to us which stories really stand out.