Sliding windows

Loading...

來自 National Research University Higher School of Economics 的課程

Deep Learning in Computer Vision

61 個評分

National Research University Higher School of Economics

Deep Learning in Computer Vision

61 個評分

課程 5（共 7 門，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.

從本節課中

Object detection

In this week, we focus on the object detection task — one of the central problems in vision. We start with recalling the conventional sliding window + classifier approach culminating in Viola-Jones detector. Tracing the development of deep convolutional detectors up until recent days, we consider R-CNN and single shot detector models. Practice includes training a face detection model using a deep convolutional neural network.

Object detection problem6:15

Sliding windows3:14

HOG-based detector2:27

Detector training3:08

Viola-Jones face detector5:41

Attentional cascades and neural networks3:21

與講師見面

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

In this video, I will talk about how we can reduce

the problem of object detection to image classification.

Image classification give us an accurate way of classifying images.

For example, you can train a pedestrian classifier to

say whether this particular image contain pedestrian or not.

But how can you say where exactly is the pedestrian in the image,

and how many pedestrians in this image?

The main technique to reduce the problem of

detection to image classification is sliding window.

Consider a fixed-size rectangular window. You've chosen correctly.

The pedestrians will occupy most of the window.

You can look through all possible regions by scanning an image left to right,

top to bottom with this window.

You classify each window using the "is pedestrian?"

classifier independently.

We mark all windows the classifier has set, yes.

There are several inherent problems of sliding windows.

How to handle the various objects sizes?

How to handle various aspect ratio of objects?

A lot of windows will have a good overlap this object.

So for each object,

we'll get multiple positive responses.

What to do if object is partially overlapped with other objects,

or its shape is purely by rectangular window?

In this last case, the classifier should be really

powerful to detect partially overlapped object,

or we have that occupy only a small portion of the window.

One way to solve the problem of object size is to use several windows of

different sizes and scan image several times with different windows.

Alternatively, we can downscale images several times and create a multi-scale pyramid.

Then we can scan all scales in window of same size.

Dependent on generalization ability of classifier,

the number of scales can greatly differ.

Until recently, there was a quite a lot of scales to achieve reliable detection.

These powerful deploy techniques,

the number of scales is reduced significantly.

If object aspect ratio can significantly differ,

like for frontal in profile views of a dog,

we need to use windows of different aspect ratio.

So, the number of images to be scanned,

equals to the product of number of scales and number of aspect ratios.

Score map for specific scale can contain multiple responses of various strength.

To obtain final detection,

we usually select points of local maxima,

this non-maximal suppression, similar to how we detect edges in image gradient maps.

In multi-scale detection, we should select 3D local maxima.

Of course, elaborate learning-based techniques exist

to perform non-maximal suppression in detectors.

Modern detectors are very powerful and can reliably detect a lot single objects.

Detector failures happens mostly for overlapping objects or for small objects.

For a conclusion, I can say this, currently,

a sliding window is the main approach to object detection.

Multiple scales and aspect ratios are handled by search windows of

different size and aspects or by image scaling and creating multi-scale image pyramid.

探索我們的目錄

免費加入並獲得個性化推薦、更新和優惠。

Coursera 致力於普及全世界最好的教育，它與全球一流大學和機構合作提供在線課程。

© 2018 Coursera Inc. 保留所有權利。

通過 App Store 下載

通過 Google Play 獲取