Deep Learning Tutorial For Video Surveillance

By Brian Karas, Published Oct 17, 2017, 11:20am EDT

Deep learning is a growing buzzword within physical security and video surveillance.

But what is 'deep learning'?

In this tutorial, we explain deep learning specifically for video surveillance covering:

  • Traditional video analytic approaches 
  • Machine learning vs deep learning
  • What makes learning 'deep'?
  • How deep learning can help analytics
  • The role of training
  • Examples of training for people and guns
  • Example of training for men vs women vs old vs young
  • Filtering alarms With deep learning
  • Training data not disclosed
  • Potential Problems Across Regions
  • Hardware Requirements
  • Evaluating Deep Learning Products

 

********

**** ******** *** ****** a *** ** ********* in *** ******** ******** for *** ********* ** finally ******* ** *******-*** analytics ********. ** ** ********** to ** *** ****** evolution ** ********** ************ or ******* ******** *** ** being ******* ** * variety ** ********** ** help ********* ****** ********** and ******** ******* ****-***** criteria *** **********. *** the ******** ** **** report, ** *** ********** deep ******** ************ ** the ******* ** ***** surveillance, ****** ********* *****.

Traditional ***** ********* **********

***** ** **** ******** use ** *********, ******* typically ****** ** **** basic *********** ** **** defined ** ****** (*.*.: people *** * ******** aspect *****, ******** *** a **** ****** *****), and ****** *********** ** each ******. *** *********** step ***** ****** ***** the ****** *** ******* were, ** **** *** system ***** **** ************* ignore ******* ***** *** ground (*****, ******, ***** moving). *********** **** ****** define **** *******, ***** so **** ******* ***** be ******* ** **** were *** ***** ** too ***** *** ************ in * ***** ****.

***** ********* *** ***** ****** that *** ** *** wrong ***** ** *** image, ** *** ***** size, *** ********* ****** could ** ******** ** determine ** ** *** an ****** ** ********. In **** *****, ******* like ****** ***** ** color ********* ***** ** used ** ********* ** an ****** *** * person, ** *******, *** example. *******, ***** ******* typically *** *** **** a *** **** * woman ** * ********** from * *******, ** they *** *** ***** much ******** ** ****** details *** ******* ******* only ** ****** ***********.

***** ******* *** **** prone ** ****************** **** objects ** *** **** the ***-*** ************, * person ******** ** *** ground ******* ****** ******* clothing ***** ** ********** as * ******* ******* of * ******, ***** their ****** ***** *** uniform ********:

**** ******** ** ********* and *********** ******* ***** on ***-********* ****** ******** is ********* ******** ** as ******* ********.

Machine ******** ** **** ********

******* ******** *** **** learning *** *******, *** they ******** ********* ********** to * *******. ******* learning **** ***-********** ************ to ***** * ******** to ********* ** ***** of ** ******, **** learning ******* ***** ****** out *************.

******* ******** *** ** ****** how ** ********* * ***** walking ** *********** ********** such ** *** ***** of *** ****** ****** be ****** **** *** width, ***** ****** ** movement ** **** *** legs **** ** **********, it ****** **** ** a *********** ********* ******* of ********, ** ****** have **** ***** *** texture ******* (************ ********), and ** *****. **** the ********* ** **** fed *****, ** **** look *** ***** **********, and ** ** ***** enough ** ****, ** will ****** *** ***** contains * ****** *******.

** **** ********, *** software ** *** ****, and **** *** **** represents *********, **** ** a *****. ** **** breaks *** **** **** into ******* **********, *** looks *** ************ ****** all (** ****) ** the **** **** ** can *** ** ***** an ************* ** *** to ********* ****** ********* of *** **** *******.

*** **** ******** ********* should **** ** **** criteria ** *** *** that ** **** ******* to **** *** ******* learning ********* *** ******** programmed ****. ** **** cases, *** **** ******** algorithm **** ** **** further, ******** ****** **** humans *** *** **** thought ** ******** ******, or **** ***** **** been **** **** ********* to ******** ******, **** as **** *** ******** relationships ** ****** *************** in ******.

What ***** ** "****"?

**** ******** ***** **** the ****** **** ******* a ****** ** ************ classification ******, *********** *******, to ***** *********. * system ******* ** ******** vehicles ** ***** ***** **** learned ** ***** ** design ******** **** **********, taillights, ******** ** ****** or *******, ****** ******, and **** *****.

**** *** ****** ** vehicles, **** ****** ***** then ******* **** ******* multiple ****** (*********** ** the ****** **** ** the ***** *****) ******* for ******** ** *** learned ***** ******* ******* manufacturers, *** ****** * decision ** ***** ***** of ******* *** **** likely ** *** *****. In **** ****, "********" has *** ******* *******, as ** ******* *** most ********, **** *** system ****** ********** *** image ******** ******** ********** with "********" *** "*******" vehicles ** * ****** extent: 

***********, **** * ******** layers *** ******** *** the ****** ** ** classified ** "****", ******* it ** *** ******** for ***** ** ** 10+ ****** ** ************** in **** ******** *******.

*** ** *** ******** ways **** ******** *** be ***********, *** ***** of *** ******** ******* is *** * **** indicator ** ******* ****** performance ** ***********. **** like *** ****** **** of * ****** ** not ** ******** *********** factor ** ********** ** image *******.

How **** ******** ***** *********

**** ******** *** * better ************* ** *** characteristics **** ****** ******* objects, ******* ** ******* on ************ ** ***** appearance. **** ***** ** better **** ** ******** objects ********** ** *********** conditions, ** **** *** object **** *** ***** any ***** *********** ** expectations.

 

********* ************ ********* ******* around ********* ** ****** at * **** ** place ***** ** ****** not ** (*.*.: *********** detection *****-*****). ** **** cases, ********* *** **** to ****** "********" ********, people ******* **** **** usually ****, ** * large ***** *** ****** in *** **** ********* suddenly (**** ** ** a ***** **********). ***** they **** ***** ****** is ** *** **** first *****, ********** ******* that * **** ** pixels ** * *****, and *** * *** (or * *****, ** a ****). ** ***** able ** ****** ********* objects ** ******** ** a *****, **** ******** helps ***** ********* ******** their ******** *** **********.

The **** ** ********

* **** ****** *******'* performance ** ***** ** the ******* *** ********* of ****** **** *** training. ******* * ****** that *** ******* ** recognize ****** **** ** showing ** ****** ** body ********, ** **********. That ****** ***** ****** fail ** ********* * majority ** ******** ******, because *** ******** **** was *** ****** ******. In ****, ** ***** perform ***** **** * manually ********** ****** ******* learning ******.

*******, **** *** ******** by ***** ***** **** with *********, ** **** of ********* ** *************** of *** ******* ******** for **************. ***** ******** images ****** ******* ***** of *** ****** **** multiple ******, *** *** cases **** ******, ** a ******* ** ***** and ********.

******* * ********** ** images, *** ******* **** with * ******** *** be **** **** *********. ******** ** *** ******** ********* used *** *** ********, as ** ******** ******** of ****** *********** **** thousands ** ********** *** sub-categories. ***"******" ************* ******** ****** *,*** sub-categories ** ******, **** classifications **** "*******" ** "executive".

***-***** ***** ********* **** the ********** ** **** much *****-******* ****** *************** than *******-******** *******. *** example, * *** *** be *** ****** ** firearms, ********** ** "**** automatic ******" *** "********" (as **** ** ******** other *************** **** *****, shotgun, ***.).

**** **** ********, ** would ** ********** ** expect *** ****** ** classify ******** ************, ***** on ***** ***************, ********* impractical *** *******-******** *******.

*********, ** ***** ***** a ****** ***** ****** of *** *** *****, young *** ***, ** build * *** ******* of *********** ****** *** age. ********* ** **** type *** ******** ******* in ****** ************ ** allow ********* ** ******* customer ******** ** ****** who **** ** **** advertisements ** ****** ********. 

Gender *** ***

*** ******* ***********, ********* in ************ *****, *** in ****** ************ ** to ****** ******** *********** based ** ***** ****** and ***. **** *** uses ******** ****, ****** images ** *** *** women, *** *** *****, that ****** *** ******** to ****** ***** *********** seen ** ****** ** what **** **** **** trained **. ******** ***** include:

Filtering ****** **** **** ********

*** ********* **** ** growing ****** ***** ************ is ***** **** ******** as * '******'. *** benefit ** **** ******** every ***** **** * surveillance ****** ***** ** very ******** *********. ******* of ***** ****, *** system ***** *** *********** video ********* ***** ** narrow **** *** ********* objects ** *******, **** forwarding **** ****** ****** matches ** **** ******** to ****** ** *** object ******* ** ********, for *******, * ******, instead ** * ***, a *********, * ****** or * *****. *** example, ******* ********* **** ******** recorder ****, ** ***** **** example **** *********:

Training **** *** *********

**** * ********* ***********, there ** ** *** for * **** ** know *** *** ****** was *******, ************* ********* will *** ******* ******** used *** ********. ***** times *** **** **** in ******** ** ********** by *** ************ *** adapted *** ***-***** ** their *******, ********* ********* this **** ***** **** other ************* ** ********* in ***** **** ** re-use *** **** **** for ***** *** ******** purposes.

Potential ******** ****** *******

**** ** ******* ** training **** *** ** *********** when ******* *** ******* based ** ****-***** ******* that *** ***** ** the ********** ******, *** differ **** ***** *** product **** ** ********. Vehicles may **** **** ********* in ***** **. ****** or ** * ******** region **. ** ********** area, *** ***** **** not *** *** ****** on *********** ** ***** region ***** **** * system ** ******. ********, a ****** ***** ** trained **** ********* ***** or European ** ******* ********. Such * ****** ******** in ******* **** ****** of ********* ********, ******, etc. ***** **** ******.

Hardware ************

**** ******** ******** ******** resources *** *** *********** elements:

  • ******** *** ****** ******* with ******
  • ********* *** ****** ******* in * ******* **** a ****** ** ********

**** ** ***** ******** rely ** *** *** of ****, **** ** particularly ********* ** *** case ** *******/*********, ** they **** *** **** traditionally ***** **** ****, meaning **** *** ******** is ********* ******** *** DNN ********* *** ******** to **** **** **** a ******** *******.

******** ** **** *************** intensive, *** ********* **** specialized **** ******** *** this **** **** *** much **** ********, *** expensive, **** **** ** used ** * ****** or ********. ********* ** the ****** *** ********** of ****** **** *** training, **** ******* *** take *****, **** ** weeks ** ********, *** usually **** ** ******** GPUs ** ********. *** training ***** ******* * model **** *** *** DNN ** *** ** classify *******, **** ***** is ********* **** ***** relative ** *** **** of *** ***** ****. Nvidia ** * ****** supplier ** **** *** this ****-*********** ******** *****.

********* *** *** ** a ****** ** ******** uses * *****-***** *** to **** **** **** suitable ** **** *******. Here, *****-**** *** ** made ******* **** *** power ***********, *** *** number ** ******* **** can ** ******** *** classified ** *** *****, or *** **** ** takes *** ****** ** classify ** ******. ********* like *****/********, *** ******* **** ******* *********** **** targeted *** ****-***** ************. ****** **** *** ****-***** ********, but **** ******** ***** expansion ***** ******** *** PC-style ********. 

Deep ******** ** ************ ********

*** **** ** **** ******** Video ************ ********* ********** ********* ** *** security ******** ******** ******** with **** ***** ** deep ******** ******** ** capabilities. ************ ***** **** facial *********** ** *********** masked *******, ** *********** entire ****** ** ****** Google-like ****** ************.

Evaluating **** ******** ********

** *** *** ********* to ****** * **** learning *******, **** ***** strongly ********* *** ******* be ****** ** * location *** *********** ** similar ** *** ******** deployment ** ********. ***** manufacturer ***** ***** **** best-case *********** *********, **** do *** ****** *** proper ************ *** ****-***** deployments. *******, ***** ***** be ********* **** * period ** *-* *****, giving *** ****** ********** time ** *** * variety ** *******, ******** performance ** ** ******** over ***/***** ********** *** across * ****** **** set **** * ***** lunch-time ******* *** ****.

Future ********* *** **** ********

********** ********, *** ***** we *** ** *** deep ******** *********** ** equivalent ** *** ***** days ** ** ******* and ******** *****. *** capabilities ********* ***** *** a ******* ***********, ** most *****, **** ******** options, *** *** ***** far **** *****. ************ are ***** ***** **** on ****, **** ********, and **** **** **** used *** ******** **** we ********** *** ******** available ***** ** **** fairly ***** **** ******** to ***** *********** ** the **** *-** *****.

***** **** ** ********* need *** ***** ********* that *** ** ********* with * **** ******** product *** ******* **** deploying * *********-********* ********. However, ***** **** * less ********* ****, ************ in ******** ***** **** search *** ******* *********, would ** **** ** exercise ******* ** ****** a ******** *** ***** in *** ********* ***** of **** ********.

Comments (26)

Bob McCarvill

10/17/17 04:13pm

Although it scares people, Deep Learning will only become a bigger and bigger market as we try to eliminate humans in the workforce. Cameras will be used to monitor assembly lines and also react when an event occurs. Similar technology can already be seen used in self-driving cars.

Avatar

Nathan Wheeler

10/17/17 04:17pm

Great tutorial.  Thanks for providing it.

Suresh Yendrapalli

Duranc | 10/17/17 07:18pm

Interesting insights differentiating machine vs deep learning.

Undisclosed Manufacturer #1

In reply to Suresh Yendrapalli | 10/17/17 08:08pm

Not sure I agree with that section of the article. From the article:

Machine learning uses pre-programmed instructions to allow a computer to recognize an image of an object, deep learning figures these things out automatically.

That to me sounds like the opposite of machine learning. From wikipedia:

 

Machine learning is a field of computer science that gives computers the ability to learn without being explicitly programmed.

According to my understanding, Deep Learning is just a special case of machine learning, one of many techniques that learn automatically by processing training data. For example cascade classifiers is a form of machine learning that also requires lots of training data, but isn't considered deep learning. I have used both techniques for detecting license plates in a scene, both require thousands of images of license plates.

Avatar

Brian Karas

IPVM | In reply to Undisclosed Manufacturer #1 | 10/17/17 08:35pm

The wikipedia article on Machine Learning focuses more on machine learning applications in understanding data than on video analytics, which is the context of this report.

This Nvidia article on AI/Machine Learning/Deep Learning is more computer-vision focused. They use the following example to show where machine learning computer vision products still required some "hand coding" in order to make the learning process more effective. This hand coding is effectively helping filter objects both in and out of consideration for analysis to get better overall results.

As it turned out, one of the very best application areas for machine learning for many years was computer vision, though it still required a great deal of hand-coding to get the job done. People would go in and write hand-coded classifiers like edge detection filters so the program could identify where an object started and stopped; shape detection to determine if it had eight sides; a classifier to recognize the letters “S-T-O-P.” From all those hand-coded classifiers they would develop algorithms to make sense of the image and “learn” to determine whether it was a stop sign. [Emphasis IPVM]

This is similar to my example in the report of giving the machine learning system some coarse parameters of humans that it can use to determine what objects in the scene should be further analyzed.

Undisclosed Manufacturer #1

In reply to Brian Karas | 10/17/17 09:02pm

But also from the nvidia article

Machine Learning at its most basic is the practice of using algorithms to parse data, learn from it, and then make a determination or prediction about something in the world. So rather than hand-coding software routines with a specific set of instructions to accomplish a particular task, the machine is “trained” using large amounts of data and algorithms that give it the ability to learn how to perform the task.

I interpret the paragraph you quoted slightly differently. I do not see it as a definition of machine learning. I see it as  an example where machine learning only takes you so far, and then in some cases required hand coded improvements to improve the accuracy. If that were a definition of machine learning, then it would be excluding other methods of automated learning, such as the Haar Cascade detection that we see used here, for example.

 

Avatar

Igor Falomkin

AxxonSoft | In reply to Undisclosed Manufacturer #1 | 10/18/17 07:13am

To solve a computer vision problem one has to create a mathematical model of the phenomenon to be analysed. For example, to recognize license plates we have to define mathematically what a license plate looks like. The main difference between machine learning in classical computer vision and artifitial neural networks is that in classical computer vision this model must be explicitly defined by a researcher and then it may be refined by machine learning (for example, the best contrast thresholds may be found automatically via machine learning). In artificial neural networks the model is defined indirectly via a training set (and in general we even don't know what features the network has choosen to solve the task).

If you are intending to deploy a deep learning product, IPVM would strongly recommend the product be tested in a location and environment as similar to the intended deployment as possible.

This is a very good advice. Because if we do not exactly know which features are used by a network then it's impossible to predict in which conditions the network will work pretty good. And be ready that the network will need an additional training on data from exact site.

Undisclosed Manufacturer #1

In reply to Igor Falomkin | 10/18/17 07:26pm

this model must be explicitly defined by a researcher and then it may be refined by machine learning

Yes, but I would argue that this model can be very general such that in practice, we software developers do not have to define mathematically what a license plate looks like. Again, I cite the example of OpenCVs cascade classifier. Feed it 2000 image of license plates to be used as training data, and it will learn with no extra effort from us to detect license plates in a scene. 

 

 

Avatar

Igor Falomkin

AxxonSoft | In reply to Undisclosed Manufacturer #1 | 10/19/17 05:33am

You are right. A model may be more or less general. The both methods use supervised learning and resemble each other. I just wanted to emphasize that using neural networks we try to pass more and more steps from a researcher to a computer. For example, in cascade classifier the basic features that the method uses are (mostly) defined by a researcher. In convolutional neural networks even these features are calculated during learning process. Of course there are still a lot of work for a researcher even with neural networks :)

Undisclosed Manufacturer #1

In reply to Igor Falomkin | 10/19/17 07:55pm

 

Yes, agreed. But it is not that I am disagreeing with Brian on the differences between deep learning and some traditional methods. Where I disagree is how the article seems to define the term Machine Learning only to mean those traditional methods and excludes all others including deep learning. This is just not correct, in this industry or any other.

I think it is important to get the terminology right. For example, a company should be entitled to use the term Machine Learning in their marketing brochures without it being redefined by the article to mean something less than it is. 

 

 

Undisclosed Manufacturer #1

In reply to Undisclosed Manufacturer #1 | 10/19/17 10:18pm

 a company should be entitled to use the term Machine Learning in their marketing brochures without it being redefined by the article to mean something less than it is.

Case in point, from this new IPVM thread, this company advertises their product as using "machine learning" but doesn't mention the term deep learning

...most advanced computer vision and machine learning ever put into a ...

but they'll be surely use deep learning in their product. 

Avatar

Skip Cusack

In reply to Undisclosed Manufacturer #1 | 10/18/17 11:21am

Brian, excellent article! I agree with this point: Machine learning implies "learning." The examples you've given to this class of analytics is what I would call "rules based," because there is no learning. Pre-defined static rules are simply applied to images or video. Unless there is a man-in-the-loop correcting false hits or validating correct hits, and this external data is used to modify or update the analytic rules, I don't see how "machine learning" can be an accurate descriptor.

Finally, thank you for not gratuitously invoking the use of "artificial intelligence."

Avatar

Brian Karas

IPVM | In reply to Skip Cusack | 10/18/17 11:35am

Machine learning, still learns, however it is "seeded" by hand-coding some basic constraints as I outlined in the article, or as Igor showed with the license plate example.

In DNN's, you feed the system a series of images, essentially telling it only what they all have in common ("these are pictures of people"), it then learns the commonalities and distinctive components on its own, without having to pre-define anything.

Undisclosed Manufacturer #1

In reply to Brian Karas | 10/18/17 08:38pm

Machine learning, still learns, however it is "seeded" by hand-coding some basic constraints as I outlined in the article

Not in practice, not always at least. The definition of machine learning is broader than that, you are confusing the limitations of some traditional machine learning techniques with the broad definition of machine learning and in doing so are excluding some machine learning methods that require no such hand coding of constraints. It would be more accurate if you said "traditional machine learning techniques versus deep learning", rather than "machine learning versus deep learning".

Igor showed with the license plate example

I would say that the mathematical model that he is referring to is not necessarily one created by the software developers who are creating an LPR system, but a general mathematical model created by the researchers/designers of the algorithms and built into the algorithm itself. These models can be general enough to detect license plates, or cars, or faces depending only on the training data. Again, I cite my experiments with the OpenCV cascade classifier, I got it to work quite nicely with LPR and no hand coded constraints or mathematical model was required.

Avatar

Brian Karas

IPVM | In reply to Undisclosed Manufacturer #1 | 10/18/17 09:16pm

There are multiple examples comparing machine learning to deep learning, where the machine learning system is given some hand-coded parameters to look for, and then uses that essentially as boundaries for its learning process to adapt its recognition of objects.

Here is one example: Deep Learning vs. Machine Learning – the essential differences you need to know! specifically when giving an example of a machine learning system designed to detect animals in images:

If we solve this as a typical machine learning problem, we will define features such as if the animal has whiskers or not, if the animal has ears & if yes, then if they are pointed. In short, we will define the facial features and let the system identify which features are more important in classifying a particular animal.

Now, deep learning takes this one step ahead. Deep learning automatically finds out the features which are important for classification, where in Machine Learning we had to manually give the features. 

Here is another example:

So what’s the difference between machine learning and deep learning then? Deep learning technically is machine learning, but while a standard machine learning model would need to be told how it should make an accurate prediction (by feeding it more data), a deep learning model is able to learn that on its own. 

And another:

The advantage that deep learning has over alternative forms of machine learning is that while the others need to analyse a predefined set of features on which they base their predictions, deep learning can identify the individual features itself.

For example, if a system wanted to identify human faces in a photo it would not need to be first be fed the individual features, such as noses and eyeballs. It could instead be fed an entire image that it can scan to understand the different features in order to make an independent prediction about the content of the images.

For the context of this report, which is video analytics/machine vision, machine learning systems more often than not involve some amount of hand-coding of parameters that the system uses to determine what sub-section of the image it should analyze and learn from. DNNs have no such hard-coding, they do not need to be told that humans tend to have vertical aspect ratios, or vehicles horizontal ones (as a very simplistic example).

Like many complex topics, there are exceptions to this, and we could write very in-depth reports on these exceptions, etc., but it would be beyond the scope of a 'tutorial', and probably more appropriate to researchers and developers.

 

Avatar

Skip Cusack

In reply to Brian Karas | 10/18/17 10:00pm

Brian, you are right to take a step back with all this. One take away from this discussion is that it reflects the industry's loose understanding, and loose vocabulary, regarding some new technologies. Adding a couple more abstraction layers to a NN, calling it "Deep Learning" and heralding it as the next greatest thing is really not very productive. It seems like every other vendor at ISC & ASIS is touting deep learning, or AI, but details are thin on exactly what they are doing and more importantly how well it really works. I'm optimistic, but cautiously so. 

Undisclosed Manufacturer #1

In reply to Brian Karas | 10/18/17 10:01pm

My original point is that it is not machine learning versus deep learning, as deep learning is a type of machine learning, rather it should be deep learning versus traditional machine learning and that the limitations you describe (requiring hand coded exceptions) are not implicit in the definition of machine learning. The titles of those articles are just adding to the confusion.

From experience, yes, if you put together an entire LPR system based on traditional machine learning techniques, you will definitely have more stages in the pipeline and more special cases that you have to write code for (e.g shadows, tow bars and rusty bolts obscuring letters, different color backplates etc) whereas the deep learning approach seems to learn everything so well that you can replace these several stages with just a single stage. The DLNs  do work really well.

 

 

Undisclosed #3

In reply to Skip Cusack | 10/19/17 02:57pm

"Finally, thank you for not gratuitously invoking the use of 'artificial intelligence.'"

Funny... cuz that's exactly what Dahua leads with in their Deep Learning youtube video...

Avatar

Brian Karas

IPVM | In reply to Undisclosed #3 | 10/19/17 03:12pm

They only mention A.I. once, at the beginning and after mentioning deep learning first. They categorize it as a segment of artificial intelligence, which is not wrong.

Undisclosed #3

In reply to Brian Karas | 10/19/17 04:00pm

"They only mention A.I. once, at the beginning..."

yes - 'leads'.

My comment was playing off of Skips tongue in cheek thanking you for not using the term.  His comedic inference being that everyone else does - which was what I was confirming. 

Avatar

Brian Karas

IPVM | In reply to Undisclosed #3 | 10/19/17 05:36pm

Difference of opinion I guess. I would call the first frames/opening of the video the "lead":

Undisclosed #3

In reply to Brian Karas | 10/19/17 05:52pm

"that's exactly what Dahua leads with in their Deep Learning youtube video..."

You seem combative.

It is labeled a Deep Learning video by the the first header - which is immediately followed by the Artificial Intelligence header (which seems to linger for a long time).

i.e. their 'Deep Learning' youtube video 'leads' with the term jokingly mentioned by Skip.

This is not an opinion to be disagreed with, instead, it is simply what is.

Avatar

Skip Cusack

10/18/17 11:48am

Hand coding basic constraints to my mind is not learning, it's an initial set of conditions that are then tested against video for compliance (to within a prescribed degree of agreement). If these initial conditions are defined by a person, and tested against video without any updating or modification, there is no learning. Machine learning, or any type of learning, needs to have a feedback (either positive or negative) to induce the learning. 

Undisclosed Integrator #2

10/18/17 03:38pm

This is quite a fun way of explaining the difference between the way CPU's work compared to GPU's 

https://youtu.be/-P28LKWTzrI

 

 

David Nguyen

IPVMU Certified | 11/01/17 04:30pm

We see this level of sophistication with technology in films, but it was surprising to learn that he have the concept under way in reality and will improve in the very near future. 

 

 

Avatar

Jared Beagley

Seagate Technology
11/01/17 06:59pm

This is a very useful tutorial in helping explain some of the general concepts of machine learning, deep learning, and AI. We recently expanded our SkyHawk surveillance drives with models called SkyHawk AI so it's very intriguing to learn more about the specifics of what this all entails.

Read this IPVM report for free.

This article is part of IPVM's 6,728 reports, 907 tests and is only available to members. To get a one-time preview of our work, enter your work email to access the full article.

Already a member? Login here | Join now
Loading Related Reports