DensePose - 3D Machine Vision

[@Wj3399]

I’m grateful to know that mega-corporations like Facebook would be willing to share their dataset which required thousands of hours of annotation work. Thanks Facebook and Siraj!

[@olfmombach260]

Though you don't know how much Facebook keeps to them

[@SirajRaval]

thanks!

[@AnArchaicSoul]

@@gkcs4636 Precisely!

[@rabiielbeji8541]

I thought that we need always stereo-vision for better 3D reconstruction, But Densepose will change my point of view, thank you Siraj

[@satibel]

if you see a picture, you can represent in your head how things are, so an AI can do the same. Of course it could be a lot better with 2 cameras, but 1 is enough.

[@MrDanielphillis]

i think approximations must be made if you only have one view...but with video containing a moving camera, you have many viewpoints to improve 3D geometry

[@SpiridonovRU]

Greetings from Russia! Very interesting

[@alialtaf3412]

This is so good for motion capture. Imagine Indie developers benefiting from this technology and then creating quality games without investing millions to produce game.. So indie developers can also create AAA quality games with the help of this tech.

[@inlustrolearningprivatelim4868]

Awesome as always Siraj☺️👍You're an inspiration

[@leowiniecki9264]

Thank you so much for this video! 3D reconstruction tech is moving so fast. That projective geometry tho... :)

[@saeedarabi5542]

Perfect man, PERFECT. Keep on awesome work.

[@AntonioKowatsch]

Well, it's been 9 months and they still haven't published the source code.

[@AmericanOtter]

This is fascinating... And terrifying

[@AhmadM-on-Google]

woah this one is really interesting thanks for explaining, i gotta present this paper :p

[@ED-TwoZeroNine]

Thanks for sharing, Sirajology :)

[@edeneden97]

Great content lately! Keep it up

[@masoudmasoumimoghaddam3832]

Siraj, You are perfect. This video sounds like transfer learning. Would you make another video explaining transfer learning.

[@rahulvanmali1406]

Buffing up! Siraj

[@michaelyadchuk4989]

OMG!!!! Using this technique, I can turn New York City into a knuckles army!!! #douknodawei

[@mmkwazir]

Great video. Would suggest changing name to reflect the true content.

[@normanheckscher]

Love ya work Siraj.

[@ashfaqmulla874]

Hey Siraj, your work is great. please make a video, which explains study plan, to study computer vision from scratch. I don't have any prior knowledge about it.

[@jayantgoel5567]

How do u manage to learn so much...plz make a vedio about you..

[@hammadshaikhha]

I think he has some videos about how he learns and reads research papers already.

[@Diggy77]

He`s an AI Robot :))

[@RAJATTHEPAGAL]

Read a few days back. Awesome stuff o.O ....

[@atrumluminarium]

Siraj, will you ever be taking a look at the Earth Mover's distance (I think the technical term is the Wasserstein distance/metric) and Barycenters applied to ML?

[@nassav3]

Hello Siraj! Nice, vid!

[@nimishverma7892]

Hey Siraj, Kinda offtopic in regard to this video. Can you do a video on how the Snapchat's CrowdSurf feature work?(It lets you watch a concert with different perspective, and the audio remains ) I have been trying to study how it works, like the root technical part. TIA

[@SirajFlorida]

Wow, that's awesome.

[@MajidAli-xu5lk]

thanks a lot, the world going so fast

[@marektluczek9531]

What's the source of 1:20 video fragment?

[@ZerofeverOfficial]

I wonder if remembering just the vertexs of things in a video instead of the whole video would be good for video compression, just rebuild the video on the end of the line (the client) i.e. (insert some algorithm remembering pertinent colors, peoples clothing and such, but not buildings, grass, trees, concrete, colors an AI would know)

[@aj-tt9qh]

Great video

[@MrChatmoon]

excellent. Thx you for sharing. I am designing, making and prototyping a kinetic sculpture of flock flying birds. I would love to find any material that can help to transform a 2d flying birg (video) into a 3d skeleton/wireframe model. Any suggestion?

[@ccerrato147]

Damn! This is IMPRESSIVE!!!

[@abdelhaksaouli8802]

funny thing, i'm working on torso detection project, i was like pose estimation state of art method give only 4 point, why i don't build something that give the whole front torso, so i started to look for data to annotate it and apply keypoint regression. and i found densepose x')

[@OwenIverson]

Systems of oppression never had it so good! This tech and the policy that will follow needs to be a major talking point with all politicians.

[@larryteslaspacexboringlawr739]

thank you for machine vision video

[@dranzerashi]

Can we use similar 3d annotations for detecting say vehicles from images as 3d objects in place of LIDAR

[@gorannovaks]

Hey there, Siraj, so when it becomes possible to capture motion from image sequence / video and transfer it into digital skeleton with all constraints, morphing, weight painting for right deformation?

[@6lack5ushi]

This is getting crazy! how similar is densepose's ROI align method to YOLO?

[@KnThSelf2ThSelfBTrue]

Is this the real Siraj?

[@myspacetimesaucegoog5632]

For speeding 3d animation this sounds like a dream come true. But then is animation itself going to rapidly lose it's charm? The power of this for Big Brother surveillance is also a double edged sword for human freedom. The importance of democratising this type of tech is immense and Siraj is doing a superb job in this regard. Phew!

[@vencendoonline]

Do you think that Raspberry Pi 4 has computational power enough to detect bodies with DensePose?

[@unboxwithaakash]

Who to follow for computer vision in twitter? Thanks for video 😄

[@crosstuck]

Thanks for being awesomeness #Tyler #TeamTyler @TeamTyl3r

[@radhikajain5098]

Can you also have some videos on openvx ?

[@thegamechanger7157]

OK the pose estimation will come. But it requires some technical systems

[@serhattural3168]

Nice share. But Machine Vision title is irrelevant

[@gshan994]

This was great. Do u have anytime worked with raspberry pi and computer vision?

[@videossoma]

I am working on NLP. I need to build an app that reads bunch of documents and answers questions related to the documents. please help how to do this. Please make a video on it

[@batosato]

Is it possible to use this method to generate 3D model of buildings?

[@gshan994]

batosato this was specifically used for body detection and 3d modelling of humans but if u use a dataset of buildings then it can create a 3d model of building

[@batosato]

Great. Is there a dataset of buildings? I reckon training the algorithm again on a single GPU would take ages!

[@MolemanAnimation]

if you wanna create 3d models from a set of images look up "photogrammetry"

[@satibel]

you should be able to get lidar datasets and combine them with a picture to generate a depth map, maybe not exactly what you want, but good enough. buildings may be harder than humans, because they have different shapes, from a skyscraper, to a castle, they change a lot more than humans who don't have more than 2 arms, 2 legs, a torso and a head. if you concentrate on skyscrapers, you can approximate them as a box, which should be quite easy to do though.

[@MrDanielphillis]

in conjunction with google street view - that should be doable for Machine Learning generated 3D models of buildings

[@drawwithpeter6950]

I may have missed this, but does this work in real-time from live video footage?

[@harshitaarora6319]

Wow, this seems like an awesome idea. One application of DensePose that comes to mind is projecting dresses, shoes or pretty much anything (converted from their 2D image, I guess) on people so they don't have to order something and try it out and return it because of size issues, among others. Could improve online shopping experience a LOT.

[@harshasrivatsa9070]

I think Amazon Echo Look has the application feature of taking pictures of a person and giving recommendations of dresses and how the person will look in it.

[@michaelbaumgarn4733]

Your story makes me so sad. I saw the post on hackernews and after a quick google search I found this collection of all the posts: knowyourmeme.com/memes/events/harshita-aroras-crypto-price-tracker-app-controversy#comments_form On the one side there is a lot of people apparently harassing you, which is not ok, but on the other hand there is also people blindly defending you despite the screenshots indicating that you did indeed lie. I found this thread to be the most unbiased about your case: www.reddit.com/r/programming/comments/7vggpq/16_year_old_app_developer_plagiarist_or_prodigy/ I would wish people could have an open investigation about this. I would be fine if it turns out you didnt lie, but so far it looks more like that you did do lie. This will have no effect on my life and on the world in general - we all will have forgotten about this in about 3 months. But I just want you to know that people like you make me sad and I think your contribution to the world is negative. Overa

[@olfmombach260]

I know some of these words

[@leander8986]

If you were completely new to machine learning and wanted to get started, how would you proceed?

[@satibel]

I'd say get a problem that can be solved using ML, and solve it. Something like an AI for sudoku, or whatever game you like could be an idea.

[@satibel]

by the way, jabrils did a great 4 episode series of videos on creating an AI using a genetic algorithm for a video game he made. kzbin.info/www/bejne/kImVeayraryifMk

[@radhikajain5098]

Awesome man !!!

[@maitreeable]

How can you make 3d character from 2d video at the first of this clip video. Thanks

[@hassanmahmood7284]

very nice . You put good efforts. I am amazed how you do this. So consistently so Good....... may God (the One who creates everything in universes, multiverses whatever where ever its, the One who is not bounded by time and space, the One who is all powerful and merciful) bless you.

[@Jacob_46]

just awesomeee

[@oscarchampion5842]

Deep fakes 2.0 is coming

[@KanalMcLP]

Oscar Champion current version is actually 2.1, so maybe 5.0

[@oscarchampion5842]

It’s just a matter of time till they add body swapping and we can swap Nicolas Cages entire body

[@satibel]

or you can animate his body however you want, pornstars will need a new job XD.

[@skotch_izolentovich]

Porn of future is coming!

[@rageshantony2182]

Sorry . I can't understand ..

[@ברקמנחם-ק3ק]

is someone know how to use it for classification of movements like standing or sitting?

[@akashnagaraj3954]

what is the papers name?

[@maroben225]

Is the source code available ?

[@caner19959595]

RoIPool or RoIAlign?

[@sniperdaoud]

you are the best ;)

[@inspectorcrud]

Would be great if we could render old movies like some like it hot into 3D

[@isacscjr]

true, and re-render it in 4k

[@w.maximilliandejohnsonbour725]

Interesting...!!!!!.

[@evanperrygiblin]

Ive been using Winmorph recently to morph faces, I know if I just wait a couple of years there will be an app that will do the whole thing for me... and my avatar

[@saminchowdhury7995]

These AI inventions always makes me demotivated. Can we still start something unique and meaningful given we live in a world where Google and Facebook comes up with groundbreaking ideas.

[@apoorvwatsky]

Samin Chowdhury I kind of feel the same dude. Siraj what do you think?

[@notaras1985]

thats bogus. true problems are food scarcity, injustice, and energetic crisis. thats what they must solve. all of us

[@MarkJay]

of course you can. Yes the AI groups at Facebook, Google, Stanford, CMU are putting out impressive stuff, but it's all open source so anyone can study their work. Yes it will take a lot of time and learning but don't be discouraged, you can add value and maybe even create your own new improved model

[@schrodingerscat3912]

DeepMind started outside of Google. You can still be creative and develop original products that will be of value whether you work for a tech giant or not. Look at Siraj, he's enlightening and inspiring people to engage and pursue this field. That's just as groundbreaking as the ideas that will come later.

[@SirajRaval]

vertical not horizontal products

[@rushikeshdarge6115]

how to execute this code ? HELP???

Wow, nice!

[@Ruhgtfo]

can make another tutorial how to make Amazon Go store alike feature computer vision example~ :D

[@md.rijoanrabbi99]

How you know all of those man?? how it possible?

[@ferdiansyahzhultriza9926]

Technology has grown up more beautiful than meg on family guy. There's another lib called openpose, did you already see that?

[@srikumarsastry7473]

Amazon GO?

[@JayRodge]

Notification Squad

[@zhongzhongclock]

Oh! It's a new monster in technology world. Lots of fake videos could be more easily created. It's good or bad? I don't know.

[@bruhaspati560]

It suraj!!

[@Dman8s]

Need more colours built into computers. Six digits hexadecimal is limited.

[@KanalMcLP]

Can't they make a break in AI inventions while I am writing a paper for school on it?

[@satibel]

you can also accept that your paper will be obsolete by the time it comes out :p

[@darogajee3286]

Have some consistency... Make some playlist of your videos....

[@zabique]

Oh, hi Raj

[@face8953]

We are the first to be combining facial recognition with fog computing and data mining. Our goal is to make security accessible and affordable. 3 main benefits of Faceter Affordable, not expensive User friendly, easy to use Using machine learning intelligence

[@RiteshKumarMaurya]

First...

[@peacemaker7953]

Dear Siraj this is very heavy for loading on mobile devices !!

[@abhijeetghodgaonkar]

Yay

[@danieli4055]

dem moves doh

[@wheelbegood]

this is all very great and I'm very interested in this kind of technology but I have a moral problem with that - because surveillance of us citizens is getting much easier for big companies and the government I don't think that this is a good evolution.. Here we have it, the old problem of the double-edged sword :/ I think in the next year(s) we should think of the ethics behind AI very carefully and we should try to maintain our privacy as a citizen

[@xuzhenlei5949]

nice!

[@anuragmishra1315]

Kewl!

[@DesiDiscovery159]

hi, i am stereo compositor, i want to know how to do this?. i dont have python knowledge

[@aswathm78]

Hell No . Its like people actually took their inspiration from Black Mirror

[@AviPars]

Epic

[@tuanva6484]

hay đó em trai

[@chrischappell7643]

ồ.... ra cao nhân là anh trai của siral à ... hè hè

[@tuanva6484]

thằng em họ mà bác :3

[@Y3nnick]

Can you ask Elon Musk if he's willing to do your next KZbin Video #DeepFakes :P

[@Krommandant]

Your green screen is bleeding onto your hands. What's up with that?

[@cooky123]

Hey Stop dude, You're uploading too much videos. You're insane!

[@rushikeshdarge6115]

please!! code is not working? can anybody help me to execute this code?

[@urgisjot]

Siraj could you do an explanation of how to use Vnect? gvv.mpi-inf.mpg.de/projects/VNect/ github.com/timctho/VNect-tensorflow

[@kaushikraghupathrunitechie]

3D Human Pose Estimation demo I made: kzbin.info/www/bejne/j6Dbf6aKjNqUfpo

[@pushpak-mahajan28]

1st

[@imshafay]

The technology is too fast mate tooo fast.

[@simpleman5098]

Can we use this technology to create a shitcoin?

[@kunal2010cool]

Difficult !