Marco Polo is a deep learning computer vision company we actually picked up one of the hardest thing to do income supervision which is fashion recognition and time if Im a female you know I picked something that is relative to my interest but actually there’s a few more reasons so fashion is actually the most needed sector in computer vision especially for visual search thank you I mean just like any other search company like Google or Baidu in the world they are a text search company and were.
basically a visual search company so there’s different tiers of visual search there’s facial recognition which I consider nowadays is not that hard to do anymore its really just a data problem how much data you have and then above that is product recognition so product like a chair or or microphone all of them have their distinct shape for for each object you have to train for that shape and then above object recognition.
Is actually deformable object, recognition, which is fashion, I was wearing the jacket. You know, as I move around the shape of the jacket change, so we actually end up picking something much harder than facial recognition harder than furniture. Recognition is also one of the most needed because its very hard to describe fashion. How do you describe this jacket on Google like its extremely hard, but its much easier, just to submit an image and have computer vision take care of it? Okay, so deep learning, computer vision had a really big breakthrough around end of 2015 and
beginning of 2016 Ive known many many startups theyre trying to do product recognition especially fashion recognition even from like 7 years ago and all of them have pivoted or felt because they did it too early like detection was not even there I actually tried to build this technology myself from scratch around John 14 and realizes its just so hard like the accuracy is low that no one would use this arent you – Im 15 and turns 16 I.
ran across a few papers published in cvpr if you guys know computer vision there’s like cvpr cumference icc VEC cv so these conferences this world scientists publish the latest findings latest research so if you follow those conferences you always find the latest publication what is in a Singh and through following those conferences Ive seen a few papers about recognizing deformable fashion objects from a very.
messy background at that time the accuracy you guys know sensed Im one of the actually there’s only two unicorn AI her company in the world one of them is cents time the other time is the other one is meltwater and theyre actually more like machine learning company theyre all theyre both like there are probably the only two billion dollar deep learning companies and same time I published a paper their accuracy was 16% that’s Mitchell on 16 and now our accuracy is 96% and you can ask me how.
do I come up with the accuracy well we we basically you have to test whether its accurate not especially subjective so we we use a very objective way so if we have the product in our database finding that product from any random image on the number one or number two spot is 96% and that’s about the highest people ever seen so so far were the most accurate fashion recognition company in.
the world and we also just developed a video recognition technology as well so why we why we focus on video so video has always been a focus it just like you have to before your finished video you have to finish image while were doing this today you know the richest company in the world right now theyre divided one of them is the content literal content platforms were talking about Netflix YouTube theyre getting all the eyeballs.
because video traffic is gonna be 80% of the Internet traffic by 2020 were all addicted and on the other side you know there’s Amazon there’s all these retail companies as you know they get the most transactions so the content get most eyeballs and the e-commerce gets most transactions but the eyeballs guys the the content ones they actually dont have much revenue like even Netflix theyre throwing away like 2 billion a year just buying contents from everywhere theyre not profitable at all like most of these content platforms theyre losing money buying content just.
there in the middle of spending money to conquer the land and for the retailers ecommerce they make a lot of they have a lot of revenue but they dont have enough eyeballs especially if you’re not Amazon everyones trying to manage their own platform even Macys here you know theyre losing their revenue projection they lost like 9 billion compared to the previous years in the last few years theyre all losing to Amazons traffic so how do we solve this problem we can.
bring content more revenue and we can bring rap ecommerce more eyeballs by making advertising more organic so this is the problem the mission of the company were basically saying on the left is annoying us on the right is organic ads that user initiatives so I like this shirt I click on it and I can buy it or even if I dont buy it the fact that I see it is value so that’s like advertising called CP n versus CP s.
so this is a quick experience I can click on a button here and the products will pop up as you seen the video I can add the product to my shopping cart and see other similar products also using computer vision and finish the checkout and thankfully today you know were in a stage where checkout technology are very events and theyre all competing with each other their strife there’s Shopify Magento so you can integrate with any one of them and get a smooth experience because content platform they dont want their traffic to leak they dont just because they.
want revenue it doesnt mean they want you to see this product while watching a TV show and then next moment you go to Amazon they want you to stay on their side so having a checkout piece is actually crucial this is another experience where you can pause the video and then we show you not only will show you where the actor is an actress you know Amazon Prime they have a feature called x-ray you pause the video youll see where the actors here is like we call the x-ray for products not only you.
see actress you can also see what theyre wearing okay so this is a quick product demonstration this is our power demo our technology work in the back earlier you guys seen the application first we started building the photo recognition and our photo recognition is using you know come by comparing the feature vectors on the left to other potted 8 abase we actually activate it they help your brands and total 5 million products from across the US you.
make a signature for each one of them like a future bacteria you can compare and do a match and also we automatically recognize the products from any given we know the location of the product on the image as well this is our Chrome extension which you can download and Shop any image you see on the Internet it doesnt matter if you’re on Google or if you’re on Pinterest and this is our video recognition its built on top of photo but its actually a few more piece for example tracking because video is basically moving.
pictures but every second lets say there’s 96 frames you cannot treat video like 96 images that means you have to search 96 times every second it will kill your system so here we added like basically similar technology like civilian camera you know we tracked the product where it goes so we know exactly in France that one 2005 is the same product were not searching it twice and a wasted trip strategically pick a few frames during that period and only search once even if the Prada disappear and reappear well still know that is.
the same product and heres how its application were actually currently working with one of the pic talks 6 year company to make a short video shopping app short video is now that the hardest thing like the short video for videos really keep your attention on your app compared to long videos because people get sync into one long video story but when you’re looking at short video you you’re never satisfied 15 seconds gone and now you’re like what next you can keep watching hundreds of.
short videos and forget about time and that’s actually according to some statistics in China tic toc is actually a now becoming one of the biggest company in China they actually are winning by do even though tic toc sounds like a commercial app but they have thousands actually have 3,000 AI computer vision scientists in the back and most of them are video recognition scientists theyre not using video recognition for shopping yet but they use it for like changing the way you look in the video making a look prettier.
and some special effects things like that hmm so I think for the biggest problem for most AI company out there is how to stay focused you know we have really smart people in our team but the smart researchers tend to say yes to everything they will say yes I can do this and yes I can do that I can do anything so when you have five clients and each one of the five clients they can you do this they want different things and most a I company even some.
most famous AI companies out there will say yes to all of them and try to monetize all of them that’s not a good way like I think weve seen failures from AI companies doing that the most important thing is to stay focused and you want to stay focused on a big market not a small market we actually try to have our image recognition just working with retailers and then we realize that.
market just too small because retailers you know they dont pay that much and then we realize you know what the rich guys are a platform content platforms theyre spending billions of dollar creating their own content why dont I give them a new way to monetize the content so this actually end up become our current focus you want to make sure you are serving the rich guys and you.
know AI company is not a charity the global market size and visual search wins I mean I think for most investors I mean were all terrible investor what they think about a company the number one thing is are we in the right sector is a sector of growing dramatically over years sometimes they might not being blessing you they might imagine in the trend actually most sophisticated master investing both the investing a person and their investing in the sector is growing really fast so the sector were in wearing the digital video market and.
digital video market every five year there’s like a new revolution and were betting on that were gonna be the next revolution it used to be pre-roll you know you can watch a hula show and then theyll stop you for ten seconds every hour and this is our current clients we have I want to bore you with this we definitely way boy is actually the Facebook of China so.
were working on making all their celebrities photo and video shoppable like me is actually the sister company from tech child that were working with were still a true big company were giving them our technology to launch so so for big did an AI company what is the most important thing is that revenue I would say most a I company take a few years to complete a product where you can launch so before that you know revenue is not sustainable you have to focus on your KPI what is your KPI for.
us our KPI is our search to purchase rate and we have never changed if you stay focused on your KPI then you can worry less about revenue so this is the especially in New York the New York investor is always asking you what is your revenue because theyre all financial background but the truth is you’re gonna be very firm and with them you say were a technology company and revenue will come if our KPI is growing at the same time you know another.
problem that most air company encounter is that how valuable is on my model everyone every scientist think the model they built is super valuable that no one else has they dont want to give out to anyone that when we work with really big companies they want us they want our model to run on their server because we dont want it its actually better for startups as well because you feel wrong the model on your server for these big companies it will kill your costs the.
AWS costs you know the cloud computing cost is huge so what we learn is you know your model is not that valuable its only a matter of time for any big companies to know the ingredients your model the value is really the data so the more youll get out of the closet the higher chance you can survive so we actually willing to put our motto on any big clients server in exchange for the data because data is our lifeblood and how do you be profitable.
well most companies are not profitable including the fortune 500 companies but you have to aim for probabilities so one of the thing I learned is like you know you have to learn what is your cost of goods sold and for us our cost of goods so for AI company is computing cost you know and also annotation cost and of course talents but you can write talents into R&D expenses because they are research handling expenses if you’re cops if your cost of goods sold is healthy that eventually youll be.
profitable so we actually learned to lower our costs by moving away from Cloud Servers I would say five years ago everybody say stupid to have your local machines and now were actually moving away from Cloud Servers just because theyre getting so expensive especially AWS I mean I know Asia will give you a free credit and Google will give you a free credit because theyre fighting for the you know for the market AWS is of vampire and we actually realized if we.
run the same thing on local machine we save the cost by we only spend about one-sixth of the same cost for that computation costs so now were actually doing the reverse strand moving everything to machines and the Challenger is a short cut of an AI company I would say the biggest challenge is infrastructure because you know we start through a demo and to build a demo you need AI scientist you need engineers but can you launch the technology was really big clients on the.
scalable infrastructure and I realize you know from being New York and China hired really good infrastructure talents for years those talents actually much harder to hire than you know a PhD a scientist because those kind of people theyre all reading big companies who have the experience of handle really scalable technology theyre in big companies getting paid at all for you to be able to steal them and say hey we dont have much traffic here to make patchouli use but we think will.
potentially have traffic in the future is not that attractive so the shortcut we find is actually try not to do everything yourself we partnered up with one of the biggest video civilian camera infrastructure company in China and they have infrastructure in every single city and we basically put our motto and wrong on their parallel computing infrastructure and that problem was solved you know I have no idea you know how do I have to spend ten million dollar to fix this so there’s a lot of thing you know try not to do alone and thank you for the time I.
think were also hiring especially infrastructure talents as you can hear from me and also AI talents the company is definitely expanding we rate previously already raised six million dollar were closing another five million right now so were looking to hire another five to seven really deep AI talents and you have to have minimum three to five years experience and that’s about it thank you.
Leave a Reply