Oversegmentation

Loading...

来自 National Research University Higher School of Economics 的课程

Deep Learning in Computer Vision

59 个评分

National Research University Higher School of Economics

Deep Learning in Computer Vision

59 个评分

课程 5（共 7 门，Specialization Advanced Machine Learning）

Deep learning added a huge boost to the already rapidly developing field of computer vision. With deep learning, a lot of new applications of computer vision techniques have been introduced and are now becoming parts of our everyday lives. These include face recognition and indexing, photo stylization or machine vision in self-driving cars. The goal of this course is to introduce students to computer vision, starting from basics and then turning to more modern deep learning models. We will cover both image and video recognition, including image classification and annotation, object recognition and image search, various object detection techniques, motion estimation, object tracking in video, human action recognition, and finally image stylization, editing and new image generation. In course project, students will learn how to build face recognition and manipulation system to understand the internal mechanics of this technology, probably the most renown and oftenly demonstrated in movies and TV-shows example of computer vision and AI.

从本节课中

Image segmentation and synthesis

In the last module of this course, we shall consider problems where the goal is to predict entire image. These are semantic image segmentation and image synthesis problems. Modern CNNs tailored for segmentation employ multiple specialised layers to allow for efficient training and inference. Lastly, we will get to know Generative Adversarial Networks — a bright new idea in machine learning, allowing to generate arbitrary realistic images.

Image segmentation3:54

Oversegmentation4:14

Deep learning models for image segmentation7:49

Human pose estimation as image segmentation8:24

与讲师见面

Anton Konushin
Senior Lecturer
HSE Faculty of Computer Science
Alexey Artemov
Senior Lecturer
HSE Faculty of Computer Science

Let us look again at the requirements for the segmentation algorithm.

Methods that more or less meet the requirements are

called oversegmentation or super pixel methods.

They're called so because we can segment a region of

one object for a larger number of fragments.

Oversegmentation is the process by which the objects being segmented from

the background are themselves segmented or fractured into subcomponents.

During years of research into the problem of image segmentation,

many methods have been developed,

starting with heuristic methods like Region growing,

which we specify a set of initial points and

gradually attached to these points neighboring pixels that have the same characteristics.

Other approaches includes split and merge,

where we divide the image into regions until

the resultant regions are homogeneous, and energy-based methods.

One formula is the segmentation task in the form of optimization of

contours and then iteratively updates these contours.

Probably, the most advanced yet relatively simple approaches are clustering-based ones.

Let us discuss a few of those in more detail.

One of the efficient oversegmentation schemes is the Mean shift algorithm.

In Mean shift, one represents every pixel in the image with a feature vector, color,

histogram of gradients and pixels of filtered image may serve as features.

The idea of this approach is to search the feature space of

both pixels for areas with large probability density mass,

most of high dimensional feature distribution.

In this slide, we display the points in the feature space and find the density maximum.

To perform the search for a local density maximum,

we examined the neighborhood of every point in

the feature space and estimate the local density via

the non-parametric estimate known from statistics such as Kernel or Parzen estimates.

We may then compute the shift,

the direction in the image where density increases in the feature space.

When the procedure is complete for all points in the sample,

we cluster points according to the simplest possible rule.

A cluster is a group of points for which

the search procedure leads to the same mode of distribution.

You can see on the slide that Mean shift divides the segments fairly well by the colors.

Having this in mind, one may note that image segmentation is

nothing more than clustering in the feature space represented in pixels.

Lots of well-known clustering algorithms are based on k-means clustering,

an approach dividing points using some kind of a distance in the feature space.

SLIC or simple linear iterative clustering is

another unsupervised segmentation method which is base on

k-means and adopted to image segmentation task.

The idea of this algorithm is to compare

the pixels that are spaced apart from each other in the image by

a distance not more than s.

The hyperparameter s can later adjust the size of super pixels.

K-means requires initial approximation of cluster centers.

In SLIC, uniformly distributed points separated by the distance s are used.

As in the k-means algorithm,

the cluster satirists are updated until the summary change in

all clusters is less than the operating threshold.

Here you can see that SLIC method outperforms other oversegmentation methods

not only by the quality but in fact also by the computation coast.

In summary, oversegmentation as an approach to

segment images iteratively by splitting them into smaller subsegments,

a lot of approaches are developed and most of them are based on k-means clustering.

That concludes this section.

探索我们的目录

免费加入并获得个性化推荐、更新和优惠。

Coursera 致力于普及全世界最好的教育，它与全球一流大学和机构合作提供在线课程。

© 2018 Coursera Inc. 保留所有权利。

通过 App Store 下载

通过 Google Play 获取