{"id":1995,"date":"2018-10-22T17:51:43","date_gmt":"2018-10-22T20:51:43","guid":{"rendered":"https:\/\/impresee.com\/?p=1995"},"modified":"2020-09-04T16:13:16","modified_gmt":"2020-09-04T20:13:16","slug":"understanding-object-detection-methods-2","status":"publish","type":"post","link":"https:\/\/impresee.com\/es\/understanding-object-detection-methods-2\/","title":{"rendered":"Entiendo m\u00e9todos de detecci\u00f3n de objetos"},"content":{"rendered":"

Object detection is a common problem when it comes to computer vision. Knowledge and understanding in this area, however, is maturing rapidly, mainly because of advances in deep-learning models. The goal of object detection is to pinpoint and classify objects of interest appearing in an image. This task is naturally more complex than straightforward image classification because there may be a range of objects we are focused on, as well as the additional requirement of needing to predict bounding boxes that will encompass the detected object. A bounding box is defined by four parameters [x, y, w, h], where the first two parameters (x,y) indicate a reference spatial position in the box, commonly the center of the box or the upper-left corner, and the last two are set for the width and height of the box, respectively. An example of this task is showed in Figure 1.<\/p>\n

\n

\"\"

Figure 1. Object detection in action.<\/p><\/div><\/figure>\n

Deep-learning based models, through convolutional neural networks, have had a positive impact on advances in this area, and this has already led to the technology being applied to industry models. In this post, we will discuss two of the main strategies for addressing object detection. The first one is a two-stage based approach mainly represented by the Faster-RCNN [1] architecture, and the second one is a one-stage approach represented by YOLO [3].<\/p>\n

Before the emergence of deep-learning, the task of object detection was addressed by means of a costly strategy known as \u201cthe sliding window<\/em>\u201d, where a rectangle with different sizes would move over the whole image trying to find relevant objects. For this purpose, a classifier for each of the classes we were interested in needed to be applied for each window. It was highly costly! Nowadays, state-of-the-art methods for object detection rely on a convolutional neural network<\/strong> which also implements a sliding window approach but in a more efficient way. Convolution layers are key!<\/p>\n

\n
\"\"

Figure 2. Scheme of a convolutional neural network [copyright Serengil].<\/p><\/div>

<\/figcaption><\/figure>\n

1. Convolutional Neural Networks<\/strong><\/p>\n

A convolutional neural network (CNN) is a special kind of neural network (see Figure 2) mainly focused on processing image data but it also includes extensions to other kinds of signals which can be thought as a 2-D grid of pixels. The term convolution <\/em>indicates that the network is composed of layers resembling the convolution<\/em> operation used frequently in image processing. Convolutional networks are the most successful examples within the connectivist<\/em> branch of the machine learning field. This can be seen as a simplified, cartoon view of brain function, sharing some properties with the primary visual cortex V1.<\/p>\n