Part One: Overview of Background and Central Similarities
With the rapid growth of multimedia data, image and video retrieval are becoming more and more important in many application fields, such as video surveillance, advertising, social media, etc. In order to improve the efficiency and accuracy of retrieval, researchers have been looking for better ways to express and match these data.
At CVPR2020, a new method called "Central Similarity" was proposed, specifically for efficient image and video retrieval. The core idea of this method is that compared to pure feature matching, using the center of the data to measure similarity will be more accurate.
But first, we need to understand what is “center”? In this context, "center" refers to the "average" position of all data points in the dataset. Imagine that you have a group of birds flying in the air, then the center position of the group of birds is the average position of all birds. Likewise, when we talk about the center of image or video data, we mean the average position of all data points in the feature space.
core of central similarity
Traditional image and video retrieval methods usually rely on the matching of local features. For example, we might extract SIFT, SURF, etc. features from the image and match them with other images in the database. Although this method works well in some cases, performance may degrade when encountering complex scenarios or large amounts of data.
In contrast, the central similarity method first calculates the center of all data points in the data set, and then uses this center to measure the similarity between the new data point and the data set. This way, we don't need to match every data point in the database, but only a "center". This greatly improves the efficiency of retrieval.
Python code implementation
To better understand center similarity, here is a simplified Python code that shows how to calculate the center of the data and how to use the center to measure similarity: