AI Product Manager Demo Project - Building a Delivery Package Detector - AI PM Community Session #7

Key Takeaways

I'm here today to present an AI project that I recently completed my goal was to use computer vision this was sort of part of a a lab that we did when I was in the [Music] cohort I'm going to get started I'm very excited today because we're going to have one of our alumni members presenting today in our community this I think this is kind of like a new Step um in our AI product management Community a new chapter in AI product management community in a sense that uh we're starting to kind of reach this point where um the quality of um some of the community members um also with the help of the cohort is reaching a level where um they can come in and like kind of present um you know what they're working on or share A New Perspective um and this is really good I think it's good for all of us and hopefully over time we'll get better um as we learn from each other so so with that said I want to pass this to Adam I'm very excited for that and Mahesh is also not joining us today um he's he's attending some I think talk in uh one of the universities in the east coast if anybody remembers which university is it MIT okay yeah yeah um so uh he's busy this weekend but uh very very excited to have Adam present today so U with that said I'm going to pass it to you Adam and like let me give you um share access as well so you can actually thank you um wait how's oh there you go okay perfect so I just gave you share access um and I'm going to put myself on mute so you can take over awesome thank you so much ban I appreciate it um what you and mahes are doing here and thank you for letting me be sort of part of the new uh evolution of uh this cohort where it you know hopefully becomes more collaborative and I'm not the you know more and more people are sharing the projects that they've worked on so thanks again for that um so like ban said my name is Adam um I'm a product manager and work mostly in the entertainment space I run a small currently running a small uh consulting firm um deploying some AI solutions for very small agencies um I'm here today to present an AI project that I recently completed Ed um my goal was to use computer vision this was sort of part of a um a lab that we did when I was in the cohort um and so the idea is using computer vision with data supplied by a home security camera to detect packages at my front door um I'm going to show you how I did it the tools I used and hopefully by the end um you'll all be able to uh you know do the same thing I'm going to walk through the tool um in a second and then before that I'll just give a brief um demo just to show you uh how it uh basically how it works um and then after I run through sort of how I set it up um we'll I'll kind of like give you a tour of the of the tool um show you how to train the model how to use your camera improving the model using inference metrics and then after that I'll just um we'll spend some time highlighting some of the challenges that I faced personally on this project um and I'll also just say in advance that if anybody works on this or is running into any problems um please reach out um it took me a really long time to figure this out I'm not a technical product manager and I felt like my devop skills would have like really come into handy come in handy here um and I definitely did not have those so was a lot of um a lot of GitHub a lot of Stack uh overflow just to try to figure out what I was doing but finally figured it out um also Chris Lincoln who is on this call definitely pestered him and Naraj a whole bunch of times um asking them for help and um so thank you guys um for helping me so let me um I'm just goingon to share just a quick video of my let's see what's the best way to do this um hold on one second let me see if I can get share sharing happening please be patient with me oops [Music] so so if I share this screen can you all see can you all see my screen all right yep with this video over here yep okay cool so this is the project basically uh this is a camera that I have at my front door this is a video of me delivering myself a package um and um you know the idea is essentially when I get boxes put in my front door um I get a little detection noted hey that you know there's a box at the front door um so I want to start by talking about um Edge Computing in AI obviously you know in case it's not obvious this is a no code solution um that uses um a Microsoft product called cust Vision um can you guys still see my screen all right okay cool I'm just GNA like tab around so if anybody can't see anything um just let me know so I just want to show you custom Vision so this is Microsoft's like vision service lets you build and deploy your own image models um out of the box um it's very kind of quick to like show how powerful this tool is um I've created some models here but there's actually some like out of thebox models just to like demonstrate it but I'll just kind of show you real quick so this is one of the models that I've created and I'll kind of go into this and you can see that all of these images um have been tagged um I you know just as an aside My Wife and Kids think I'm absolutely crazy because I taped up a bunch of empty boxes to trade this model and I have about 500 uh I have a bunch I have about 500 photos of boxes here that I Ed to train so um that was that was definitely a challenge in some way um and I'll just kind of show you a quick test um it's super this is super easy to use um but if you kind of just like I have a just like a whole bunch of um photos of boxes in here and just to like run some inference really quick and you can see that it very quickly recognized this inde a box you can change your um threshold hold value um to like pick up uh various things if you want to get your uh model to be a little bit more precise um but very very easy great really great tool by Microsoft um and it's easy to like sort of um you know like plug plug and play so to speak um so the final project uses an open source project that's leveraging custom Vision called Vision on edge um hold on one second sorry um and vision on edge is actually a project an open source project that Mahesh uh created um and it us utilizes Edge Computing and Azure iot Hub to bring a customers's vision project um to production using their own data and so that's the idea with this project I'm using my own data using my camera at my front door and processing it on the edge quote unquote um and just real quick um you know I just want to talk real quick about like why AI on edge like why is that important um and sort of what are the general challenges and I'll talk about my own sort of you know how my own personal challenge has fall into this a little later but you know I think the big benefits obviously are going to be um you know real-time processing you have lower latency that means um you know faster potentially safer depending on what kind of um solution you're employing um for scaled solution that could be a higher you know higher accuracy for decision making um in like a factory setting which I think is what a lot of people are utilizing this tool for um they're resource efficient by Design um they're effic because there's limited capability um on the edge for devices in compared to the cloud that's not always the case um but oftentimes it's the case um they're scalable so scaling via adding more devices increases your data collection um there's great redundancy so you can provide like failover capabilities in case devices uh fail or servers fail um you know they're cost effective potentially um although there is sort of a trade-off there so while you save on bandwidth to the cloud uh there may be like additional costs associated with deploying um Hardware on on edge so it's a little bit of like horse trading I don't know if any I'm sure some of you have experience with this so it really depends on the application and you just kind of have to decide on what's the best thing for your product um I would say some of the biggest challenges so data quality um this was certainly a problem that I had so you know especially if you have cameras that are like or sensors in general that are just like noisy or just don't have like great um Quality you could get like inconsistent data um which means like your your the quality of your model is you know it's less reliable it's less accurate um security and privacy are are things that you know again it's a little bit of horse trading um it does sort of remove Transit trans transferring data to the cloud but it also adds in some insecure points um where people can actually have physical access to your devices on the edge so again it really depends on the application um you know uh model optimization is the next thing um so due to the limitations with compute or memory or storage on the edge um models you know that are built for the cloud need to be reduced um with some kind of reduction meth uh methodology like um quantization which I I don't know if everyone in the group's talked about but you know it's just a very interesting um way to reduce the uh size of models for like Edge Computing I would say that um you know while this again this could restrict the complexity of models but there's been some very good advances in model reduction like a lot of really interesting techniques so you know it's it's becoming less prevalent I would say and again you know if anybody um has personal experience with this I'd love to hear more about it um and then scalability I think you know it's again some horse trading here but like while scale it could be simpler to scale but you're also talking about managing a bunch of devices on the edge and they might not be consistent they may not be the same device you may be getting various types of um you know you might be getting different types of signal um scaled in different ways so it's you know it really depends on the solution and I think this is another re uh area where there's been great strides uh I don't know if I have the ability to call on someone but suit it looks like hi a quick question a quick question you mentioned something about uh model reduction could you double click on that please sorry say that again uh you mentioned about model reduction yeah could you elaborate a little bit more on that yeah so um you know you have we have these huge models right who that are generally built to work on the cloud um and if you're building a if you're building out a solution where the processing power of your model is actually happening on an edge device like a devkit like a Nvidia or an I BM you know smaller device you just generally have less compute Less storage less memory and so your model has to be um reduced in size in some way to work effectively to still be to have an acceptable amount of accuracy so again when you're sort of um trying to solve for your particular problem um you really need to like make these considerations like what is the level of accuracy that we need um are there you know reduction methods that will be appropriate for the solution you know how much uh reduction is acceptable in order for us to still um like maintain the type of um inference that we're you know that we need in order to like be successful does that make sense uh yes um but do we have any uh any material on this and yeah I I I'll take a look I think there was a session where mahes talked about um the like the compact models and like why sometimes you try to like basically enable a model to work like locally on a device um I think it was either on week one or week two but um let me look and get back to you um there might be a lesson on this I'm trying to kind of like go back off of memory but um we'll check and let you know if anybody else remembers I believe in some of the reading material there was some some document okay great yeah thank you sir but I I I do think uh it's you know just kind of toying around with um Vision on edge gives you a a good sense especially when you start you know like do running your own inference um definitely have a sense like you know you there is some scaling that's happening and um you know like in what real application like I'm just doing this for fun um and whether the boxes you know are recognized by my my doorbell security camera doesn't really matter but like for your solution you know how important is it and what kind of accuracy do you need so um I'll just keep going um I want to talk about the tools uh real quick that we use so obviously um I told you that this is a custom Vision this is an Azure Microsoft Azure product so you needed Azure and and custom Vision uh subscription in order for this to work um and you also need a camera um so I'm using can everybody see my uh my screen if I if I go like this I'm sure many of you know what this camera is um this is a wise cam it's like a consumer uh security camera um I do not Advocate using these if you do do this project um it's a big pain in the butt because these cameras are not very friendly to being um used in any other way but uh connected directly to the software that um wise wants you to use and what you really need to do is utilize a camera that has an rtsp feed um I actually have I'll just uh recommend one and Chris if he's on this call I think this is the one you used Chris if you're still there um this is a this is like a really I mean they're both affordable cameras and if you know I just had this at home so it was I just didn't want to purchase another camera but this is just all I had um and it caused a lot of pain and agida um because it was not friendly to to me accessing an rtsp Port um I needed a to like revert to a very old firmware that wasn't supported by the company and I used like internet archive to find that firmware and it was a huge pain um you know but uh it ended up working out if you do take on this project I would highly suggest this camera it's very affordable and uh it freely offers um the ability to access that rtsp Port which is important um the last tool that I would say you need is um a single Wi-Fi router and the reason I even mentioned that is because I was running on a mesh Network um in my house and that actually caused a lot of issues because I think it something about like which access point the camera versus my router was connecting to is giving me some trouble in terms of um port forwarding which I'll talk about in a little bit um I saw someone had their hand up I yeah I had few questions first of all yeah like why do you think that rtsp is important for CV use cases so I can't speak to all use cases but in terms of um The Vision on Edge product um the way that the data is accessed from from the camera is via this rtsp feed so this is just um you know what Vision on edge the product that I used out of the box needed in order to collect data from my camera got it so that answerers the first question second is like I saw you used wise uh yeah uh as you know that like one is like training it and then putting in like for work like I already have ring camera I have Lorex have you looked at other cameras also or like do would you are getting the same issues like with way you have to use their software only like yeah so the biggest thing is you need to have access to create an rtsp feed so any camera that has an rtsp feed will work um wise had opened that as a as a feature or like part of the functionality that they had several years ago and then they decided to hide it um and then they decided to remove the firmware from their websites so I had to again I had to revert to an old firmware in order to access that um and wise is constantly trying to like have me upgrade my firmware um so it's just it's kind of like if you get too fancy of a camera um that doesn't readily have like advertised rtsp feed um it could be a little bit annoying like the trouble that I had with it but I'll walk you through and again I think if you um as long as you look for a camera that has um that says rtsp feed which this one does I'm not really sure I know it's here somewhere but um oh here it is right at the top yeah so that's really the most important thing and I think these cameras are mainly designed to integrate with like third-party software versus using their own software like a home network versus like part of what wise is selling not to like get too tangent here but I think part of what wise is selling is like the software so that it you know does some AI tip uh trick trickery for you um but in this case it was sort of like working against me to have that so um I would just suggest like a less bells and whistles camera that has the rtsp feed I can't speak to the lorx and I can't speak to the ring but I would just check the um documentation and see if they have access to those feeds perfect thank you so much yeah you got it um all right so the first step you got to set up your camera so like I said um I had uh to run around to find this like old firmware um it was a little bit annoying to find eventually I did find it um once um I did that I was sort of able to um like access the um feed so you just kind of go into the camera um in your software for if you're using a wise camera and you go into the settings um and once you're in your settings um you can scroll all the way down and there'll be this like rst rtsp feed option and basically it generates um with a you know with a private uh with the camera's private IP um it also tacks on this like uh password and username which I was not particularly creative here um this is like very little security and you are eventually opening your um camera up to the Internet so you know I think uh this is just something to consider I was less concerned about it just because this is outside my house and I'm not really like keeping this on all the time um but if this is a like more s you know like long-term solution I would say um this is definitely like a security weak point here um because if anybody finds out um you know what the password is it's you know which I think is probably relatively easy to find um you could you know have anybody accessing this camera um so you know once you have your rtsp feed set up um you need to basically open your camera out to your um you know to the internet uh to the like Wan um so we need to do something called port forwarding so you uh basically like go into your network settings um there should be a setting here called port forwarding I think that is where I was and um you know all all cameras have like a different uh Port that they're using I think Mo a lot of them use 554 but you basically are just opening um a port um with that uh with that IP um on that port and hitting save and then you've basically done the work and you can test it there's some websites that you can use to test to make sure that the camera's open um and what you would do to test it is you take that uh link I just sent uh the rtsp feed and you replace this um IP with your public IP um and that will open through um oh then you got to add this 554 here and then it will open the camera feed and you'll be able to access the camera anywhere so that takes care of that does any I know that was a little weird I don't know how many other people have had experience port forwarding if you you guys are into like home security I feel like a lot of people who do home security think that's no big deal but just something that I had to kind of learn about and I know Chris also had similar problems when he was doing this um so after that um once you have the camera set up you basically are ready to um start setting up Vision on edge um you so again you need your Azure subscription you need your uh um custom Vision subscription um and assuming you have those accounts um you got to set up um the machine that you're going to deploy Vision on edge on so this is made to work on an edge device um and you need to um use IO uh azure's like iot Edge service in order to do that so in order to do that you have to first um set up iot Edge service on a vir virtual machine um there is a GitHub that I can link everyone to I think Mahesh and ban have links to um to like walk you through how to do that um I have it right here um but basically you're deploying um your Edge module to um your device so you can definitely use like like I said was before you can use like the Nvidia Jetson like another dev kit to do this um I did not have access to a dev kit I didn't really feel like purchasing one so I just decided that I was going to use this like virtual uh like a Linux virtual machine that you can get easily through Azure um so once you do that um and I'm not going to go through all of this but it's it's pretty straightforward um and I think I did this via like the azure shell um but it was very easy so once you do this you need to open a network Port um to access the virtual machine so this is my Azure um account and basically what you need to do here is go to the virtual machine um which is this one that I have here and access the networking settings and then again just like you did with the um your um local network you basically have to open a port so that you can access this virtual machine via the worldwide web um so I'm opening Port 8181 um that's the port that this virtual machine works on and that's pretty much it after that um you um have to in so then after that you install Vision on edge again this is like some more uh shell installation here um and I'm I'm not going to go through all of this but it's like pretty it's pretty straightforward these instructions are very easy um and once you're finished um you basically waiting uh it says you know 10 minutes but I think it took an hour for like all of uh like my runtime status to be positive and that definitely threw me off a little bit um but I'll kind of just show you what that looks like so if you go into your iot Hub and you go to iot Edge which is the configuration and basically you want your runtime response to look like this um and as long as you have everything running which I do that basically means you're set up so once you do that you go back to your virtual machine and you'll get this public IP address here and the way you access Vision on Edge is you um use this IP address with the 8181 uh Port which is here and I'm pretty sure all of you could in theory access this but please don't right now because it'll probably uh time out if we're all on it um but this is basically my live solution up and running um and I'll just give it a little second for my camera to load there you go this is really my front door right now um I get this box once a week and this is empty and the wind's blowing really hard today so this has blown down the street several times today um so this is the this is the live um solution um you can see this is now running inference um and you can see the you know based on the model that I've trained thinks this box is you know this it it thinks that you know 94% confidence that this is a box and this is a little smaller um and so this is this is Vision on edge this is what U mesh has uh built it's a really cool very cool tool and um I'll kind of walk you through all the features here and you know please stop me if you have any questions so when you come here the first thing you're going to do is you're going to go and set up a model so you um there's a whole bunch of like out of the box um models that you can get as well but I set up a brand new one you can name it and tell it what kind of objects you want to detect once you do that um you basically are ready to add images to your model and this is pretty similar to what you have all seen um there's sort of two sections here one is like all my tagged images um these actually even though I have this model trained on like 500 images of boxes um those are all not showing up here at the moment because I've redeployed this model a couple of times but they're on custom vision and um really what's coming over here is model and not the original images that I used to train this um but I do have some untagged images here and I'll we'll go back to this um uh actually we can do that maybe it makes sense to do that right now so um here's where you can upload images um these images actually are coming from inference but let's pretend like I just took some pictures of boxes and uploaded it um you can come in here and you can see that inference says that that you know it thinks this is a box with 89.7 n% confidence it does not think this is a box um but you can kind of go in here and just like draw um a square around what is a box hold on I'm just moving something around here and then you just go to the next image um so that's the same image and you can kind of just keep going for all your images I'm going to delete a couple of these you can see how like there is some sort of like color change in this camera and um you know I definitely get different kind of lighting uh depending on the type the time of day and I think working this um solution in a factory where you're presuming that the lighting doesn't change you're going to get a lot better uh like General accuracy over time um I'm going to just delete a couple of these because they're all the same um basically you can set sort of like a threshold um and so that's basically how you train your model you just kind of go through once you have your model um once you've tagged your model with with uh you know images based on the data that you've collected um you can train it here um this will retrain the model to include you know the the newest three tagged images I'm not going to hit that because sometimes that takes some time but you would hit that and then it'll say model is successfully trained great and then you just deploy it so you go over here um you uh it gives you some options here uh for deployment um you can select which model you want um you can select which camera I think you can select multiple cameras if you'd like and then there's some ADV advanced settings um I don't have this set up this Cloud messages um but you can actually send like video data to the cloud um if you're looking to do like some video analysis um I was just trying to skimp on the cost for the purposes of this demo so I did not enable the video analysis and then in terms of retraining basically you can set um like what percentage confidence do you want um to to have images being saved I have between 74 and 90 um you know if you really want to hone this in um you can kind of adjust this to your liking you can also mess with your cameras frames per second um sometimes this tool gives you a recommendation in terms of what frames per second is going to work the best um just so you know I'm I am using CPU here to run this so this is not the most efficient way to run um this Tool uh but that is just all I had access to um so I would imagine if you're running with a little bit more juice you could probably up this and run it a little faster but I know it works and functions well even even if this is all the way down at one frames per second so um it is a pretty flexible solution I would say um once you have your model deployed um you know you can be and You Begin inference you can you know you start to see these square boxes and then you get these metrics telling you what the success rate um of the model recognize boxes in the current deployment and then um as this is happening it is um capturing taking snapshots of the image and then you can go back and retrain the model by tagging these images that um the model has where theel model has inferred um that you have a box um challenges here like I said the wise camera I was being lazy I think or maybe I just didn't want to spend 30 five bucks on a new camera but um it was a bit of a pain just getting that working um port forwarding on a mesh Network um I had to kind of just shut my mesh Network down at home um I didn't really have an any experience using like the Azure cloud services personally I'm I'm not a devops guy I was not familiar with like the shell installer um it is pretty easy it definitely just slowed me down a little bit the other I think the biggest thing probably is once I got to training um this does this this tool does say that like I think 15 to 20 images you'll start being able to run inference but I needed a lot more images to get this to accurately detect boxes um if you I don't actually have any examples here but it often like I have a squirrel that comes here and it often thinks a squirrel is a box um and that really annoys me because of how many um images I have of boxes uh but you know I think a big part of this is that this camera is facing south I'm in the northern hemisphere so I'm getting tons of of lighting changes throughout the day um in the afternoon the sun is right here in the sky so I don't really think that plays well for um the quality of data that I'm capturing and you know the best days for me to Tinker with this have been like overcast days or really early in the morning before the Sun's Up or very late um where the sun's gone down behind the um trees on the right side um you know I think for me the next step here would be um to um continue this lab and um you know program um it to send me a text message saying I've gotten a box delivered um I also saw this morning actually there's a there was a like a farmer in the UK who did this um to detect rodents that were infesting his far his Gardens in his farm um so he has them like trained on images of and it was it was the this exact solution um and he was he was training it on pictures of like bunnies and squirrels and stuff and every time uh it hit a certain like confidence level he had this like um some sort of like Hypersonic frequency alarm scare them away so I think there's some like really interesting things that I've seen people do with this these types of like computer vision um and it's you know again I am a non-technical product manager the fact that I was able to like hobble this together I think means that anybody here could pretty easily do it and um I think that's pretty much it um does anybody have any questions thoughts feedback anybody interested one thing I can tell you um from my experience also um just using the custom Vision part like not so much on um you know using the video but um one product that we've also built at PM exercises some of you might be familiar with it it's called Product monkey. a that basically takes um uh product design files and converts them into engineering task tickets or product requirements um one of the challenges that we faced was how there are so many different edge cases um with people uploading files like for example sometimes um the product design files include a model um like there's a lot of kind of vagueness in the background and then there's a kind of a box that is really what's important about that page um so we actually have to use custom vision for the purpose of for example um detecting whether or not what the user has uploaded was a model um then we had to use um custom Vision uh to better determine if what the user has uploaded was a form um and then there's like so many different edge cases like we have a whole logic in the background um that uh determines uh what the right set of actions are for creating quality tickets and prds based on the the result of what we feed into custom Vision so um this is kind of an example of this I think case that we've actually been able to kind of find custom Vision powerful and in our case um we could actually get to high quality within like 30 to 40 image uploads so um it's usually like I would say 90% of the time it's accurate but sometimes the users also like upload like really difficult to detect um images um and in those cases you know it's hard for custom Vision to detect it but um I thought that's an interesting use case I a few people yeah thank you for that I was just to add to that I will say that um I don't have this on this iteration of the model but um if you use like a multiclass for recognizing it things um using your camera the model actually runs better better so like for example you know there's like this this car that pulls out on the driveway across the street from me that always picks up as a box but like if I you know I tried to train cars as like a secondary object and the model works so much better if you can just classify a lot of different things that you expect to see in your data um and then you can filter in and out the things that you are interested in um at least with this particular use case I found that to be the case so just to add to what you were saying ban yeah yeah um um I can see that Zach and chaali posted a question for you as well I think they're the same oh um I can it says have you tried training with boxes of images of different times of the day and night yes I have um definitely um I've tried training um at all different parts of the day um at first I actually only had Windows to train at the worst time for lighting um and I was struggling there and then I realized that the lighting was probably a major factor so I started um training at different types of at uh using data from different parts of the day so I was able to just capture images you know from from days where it was really sunny from days where it was rainy um we have a porch light so I was also using images with the porch light on at night um and that really really helped um the uh model's ability to like have high you know predict predictability on inference for sure one of the feutures that I like about um I think it was custom Vision if if the lad is on a call he might know better uh but like also once it gives you um its inference if you don't disagree you can actually um say to it uh look for example it was a box or it wasn't a box and then use that as training and it basically improves the quality so um I think in a real product management environment maybe what it turns out to and that's one thing that we're doing at product FY um is that on a weekly basis we look at all the inferences and then sometimes if we see that um it made the wrong call we correct it and then we asked it to do a um retraining model and that hopefully over time kind of slowly improves the quality as it covers more and more edge cases yeah I wish I could show you that in real time but often there are things on my porch that gets P get picked up in inference that are not boxes so what happens is you know it captures those and you can go in here and just like you would say you know to your point you can say this is a box so you can go like that and let's say like sometimes you know this house gets detected as a box you could delete that and then now this image is been tagged you can go back and retrain the model and it you know very helpful in improving your um results uh I have a question hi am chali hi yeah this is pretty interesting and I have been working with a few visioni products uh specifically image recognition kind of things and made them Services uh for machine parts and those kind of things uh so the question is how you handle the data drift here because often for the boxes I'm not sure about the use case but then there are always chances of the data getting drifted so how often you uh have been retraining uh the models so I haven't had them yeah so I don't really need to worry I haven't had to worry about drift because um I just haven't had the model um running inference for that long and I really don't have that many like my my d set really not that large um but I was thinking about that um I think if anything I could probably use a lot more data um and yeah drift is not been a an issue for me so far thankfully um but I but I am curious has that been uh something you've encountered and were there ways that you you mitigated the drift yeah so uh so we have encountered couple of times and what happens is uh we were dealing with vehicle parts so vehicle parts when they are being taken from different um Vehicles specifically even the same make and depending on different kind of condition they are into uh there were chances that they may have slight rust or uh some kind of deviations or wor out so those kind of things used to cause a lot of data drift so we had to go and we have to like um each time we uh we had to go and retrain it when the certain Precision was not achieved uh but then we had to make a pipeline for that so that was quite tedious uh initially but then yeah that the process when it became streamlined with the pipeline then the overall inferencing also improved and the data drift has been taken care of because in production it becomes very important right uh you have to have a regular Cadence of retraining them uh in case there is this kind of situations yeah absolutely and I'm I'm definitely doing some retraining but you know I think if my classifier was more something more specific like a specific type of box like only boxes coming from you know um like Amazon or something like that um I could see there being more you know it it encountering drift a lot quicker but I haven't I haven't seen that as a problem yet and I think it's just because boxes probably you know from a computer vision standpoint is probably pretty simple um and not really a lot of like um like data requirements to recognize you know the box versus a different shape but yeah once you get into like more specific use cases I I I could definitely see that that's really interesting thank you thank you any other questions can't see the chat so I anybody I don't know if there's anything in the chat but um I hav seen anything new being posted um here it is yeah I don't see anything new but um this was great thank you so much everyone uh for joining and thank you Adam for presenting I think this was a really interesting like I think the custom Vision on the surface it might seem like a very simple tool but it's so powerful it's just amazing how much we can do with it I totally agree I I I had the same reaction I thought it was a very straightforward tool and really if you go play around with custom Vision you can get so quickly like without even creating your own model just pulling up some of the demos um it's incredibly powerful the things that you can like the potential the to the tool has and it's really it's like in a couple clicks you can kind of see right out the box um and I really like the vision on edge concept I would love to in the future um deploy this like on a dev kit um because I think that would just be you know for no reason except for I think it would be cool um but um you know it it works well and uh you know it's there just so many possibilities now that I have it deployed and reading about the way other people have used this tool or similar tools it's it's it's pretty powerful yeah and it's interesting how there's like some jobs like Radiologists that get paid so well for um the the whole job of looking at an image and understanding what's going on in there it'd be interesting to see how um this will impact those kinds of jobs over the next few years really powerful a are one of the difficult sorry to interrupt no for sure go ahead images are one of the difficult most so I have been volunteering helping uh Healthcare organization building uh some um osteoarthritis detection kind of models and the problem what we faced is uh this the similar thing what Adam was explaining right if the x-rays were not being taken at certain lighting conditions and a little bit of glare also even we uh have all those image processing techniques applied on them still they are not precise and the and the um impact of not being precise is quite huge so it becomes inherently difficult not involving a radiologist and someone like uh anme still the human intervention is needed we are fine tuning it yeah and the amount of data that you need for like lighting variance is so much more significant I can at least at least for the application that I was using but I can imagine with x-ray yeah because you need extremely high Precision right yeah that makes sense great well this was a great session thank you so much um thanks so much um Adam and U people that actively participate that in sharing their opinion uh chali this is great sujit and kid um and anybody else who um participated by sharing their opinion this is great we're going to do more of this in case you guys are interested in also sharing anything with the community uh feel free to reach out to me and uh we can work through it um otherwise I think next week Mahesh will be back um anything else you want to add Adam before we end the session no I just want to thank you guys again for letting me present and you know I I'm very excited about this cohort turning into a more collaborative let's see what everyone's working on let's kind of talk about it kind of opening the floor so to speak um you know mahash has really like and you have unlocked some amazing knowledge and uh I love the idea of just like taking it to the next level and getting our hands on some cool projects uh together as a group so thanks thanks for letting participate of course of course and uh I've also shared the link to our upcoming cohort that starts in November 7 I believe that's the starting date we're going to have both morning and evening sessions so um the cohort is going to get larger we're going to have more alumni people so uh feel free to apply um otherwise enjoy the rest of your weekend and we'll see you next Saturday have a good one bye bye everybody