Best computer vision competitions on Kaggle (for beginners)

hello everyone and welcome to my youtube channel in this video today i'm going to talk about some computer vision competitions on kaggle for beginners in my previous video i talked about natural language processing competitions on kaggle which i found were the best ones to learn from and these are the computer vision problems that you might like if you're beginning in computer vision if you're if you have done some theory in computer vision and want to apply that knowledge to some competitions or some some real world problems so this is my list of best competition i hope you like it so the first competition that we are going to look at today is amnest no sorry it's not amnest everyone is doing amnest and uh it is it is a very simple and easy data set i you you can do a lot of things with mnist just like in my book i present mnist as a unsupervised problem so you can try things like that if you want but if you want to just do uh classification or regression kind of problems then probably you should start uh you can you can start with mnist and then you should move to some more difficult problems so this is not a very difficult problem rather it's it's quite easy it's about dogs versus cats so you are given an image of a dog or image of cat and you have to build a computer vision model a deep learning model maybe to distinguish between these two different kinds of images so it's a very simple binary classification problem you have you have a lot of data here and these are nothing but images so dogs are represented by one and cats are represented by zero and you have 25 000 images so quite simple data set maybe you can create a model in kaggle kernels without even looking at what people have done so that's what you should try to do simple simple competition simple data set to start with and after that we go to multi-class classification multi-class classification so it's flower classification and why do i include this it's because there's a lot to learn from this competition so you have 104 different types of flowers so these are flower images and you have to classify them into 104 categories and why this competition this competition i have included because it has a lot of data and when you have a lot of data you might want to use tpus or you want to use much powerful gpus and kaggle gives you tpus for free for 30 hours a week so in this competition you can learn how to use gpus for classification problems or for image classification problems or any kind of deep learning problem so you can try to use tpus with pytorch you can try to use tpus with tensorflow um the best part of this dataset is they have also provided images in tf records format for different sizes so as you can see like you have to record 192 192 to 24 to network so on so you have different sizes and they're already in tf records tf records are needed because you if you want to run on tpus it should be in tf record format and you can also try some by torch based model so if i go to notebooks i will see there are many good notebooks i'm just going to sort by most words so there can be some hidden gems and you can try to find out and uh there are many many good notebooks here and you you can try uh to take a look at those notebooks how they approach pro of this kind of problem when you have to do it on tpus one more thing that you're going to learn here is many different types of augmentations so that's something definitely you should take a look at so you begin began with dogs and cats or amnest kind of problems and then you try to apply different kinds of augmentations so you increase your knowledge about that the next competition that we are going to take a look at is c discounts image classification challenge so this is also an image classification challenge again and here i have included this because the data set is huge you the size is 58 gigabytes and it's in a very different kind of format so it's in a bson format binary json so you have to extract the data which is a binary string so so the picture or the image is a binary string in this bson file and you have to create a model based on that so if you look at the notebooks um it's it's an old competition so you don't have it's not very old but you don't have a lot of notebooks with many words but you can see you have some bass lines but here uh walter has written a notebook on processing bs and files so you can take a look at that and you can start from there but uh i would suggest you not to look at the models but this is this is just the data munching so you have different images of different sizes and you start from there you say you have already extracted the data set maybe you can create a data set out of these images and then uh do your machine learning or deep learning part of it so so far you have docs and catch you have flower classification you have a c discount image classification challenge all these are different types of image classification problems so multi-class binary classification you also have other datasets like mnist so that's something you should start with and now we go to medical images so this is a very fun challenge and it's about pneumonia detection so if if you look at the data sets now the data set is again in a very different format so the data set is dicom images and you will be working with dicom images so now that there's a lot of things that you can do with dicom images dicom images also come with a lot of metadata itself so you can probably use that metadata too and most simple way to start with this kind of challenge will be to convert the dichomom images to png or jpeg formats so uh you have uh you can do simple pneumonia or no pneumonia or you can do uh bounding boxes so we will come to bounty boxes later maybe you can just do pneumonia or not pneumonia in this kind of challenge so where you have the bounding box you have pneumonia if you don't have it it's not there so um so you can just use this target so you just have the image and you have the target and you can build a model to classify whether there is pneumonia or not and you will learn how to process tycom images how to work with medical images how what kind of augmentations do medical images require so all these kind of things you will learn in this competition there's a lot of notebooks that has been shared a lot of good ones so you can also take look look at that and get inspired from those notebooks but if you have to start then as always start on your own and then look at the notebooks on seeing how you can improve the results the next competition that we're going to look at is also rather a simple one it's about facial keypoint detection so with every image you are given a bunch of values a bunch of targets so it's a regression problem and you have these different different coordinates for the 15 key points and uh you have 7000 approximately 7000 training images and 18783 test images so in this competition you will learn how to uh detect key points on facial images and this is this is also a very interesting and very nice competition let me check if there is some notebook where i can show you some images or maybe maybe i cannot show you images here but like you look at the data so this is what the data looks like and you have images like this and that on that you have the key points represented so i don't i don't think this kernel has the key points it's more of a training oh yeah it is it has another image so it maybe has no it doesn't have any key points but yeah you get the idea so you have uh um you have to build a model you have to build a convolutional neural network or yeah any kind of deep learning model maybe are simple models i don't know you can use keras you can use pie dots you can use whatever you want to use and yeah so these are some of the images where you have like the detected key points and i'm not sure if they're coming from the model that this person has trained or these are the original ones but i see like there's some some something wrong then maybe uh it's from the model so you have you have the facial images and you have different key points and you have to detect them using building some deep learning model it's very nice very interesting challenge the next compilation that we go to is identification of endangered right whales from aerial photographs so this is this is very interesting interesting problem these are pictures of right wheels these are kind of whales and uh you have to identify them so if there are right builds or not if i'm correct so let's look at the data so you're given aerial images and each image contains a single right wheel and you have to do a face recognition system for whales so you have to identify the faces of the wheels so um quite interesting problem so what is the like you have the image and you have id for whales so different types of whales so uh in this competition you will also learn a lot of things maybe you can you can create a face recognition system for whales and so you what what you can do is you can create a bounding box and extract this part of whales and uh then then create your model uh so if you look at notebooks and discussion this this is a very interesting very nice challenge um no notebooks okay maybe it's because it's so quite old but in discussions you have a lot of good discussions and the first place solution is also there which is quite amazing and you should definitely take a look at that now we move to uh so these till now where we have looked at image classification challenge we have also looked at uh pneumonia detection challenge which was detection of bounding boxes but it is also a classification challenge so we we can do the classification challenge first and then go to bounding boxes but anyways the next challenge that we have is tgs salt identification challenge in this challenge you are given a data set so these are seismic images and your your challenge is to build a segmentation model to identify where the deposits are where large salt deposits are so this this is a nice challenge to learn about image segmentation so once you're done with classification and regression kind of problems with computer vision or images you can you can come to this competition to learn about same image segmentation and you see the images are not very large it's not a a lot of data only 450 60 megabytes of data so a few thousands images uh not much and you have uh i hope you have some kernels here so like you can learn about image segmentation using units uh so i think the units work quite well in this competition and you can you can learn about it uh and um yeah this is quite good competition so you have to segment the images so image segmentation is also always like a very fun problem and it's also used in medical imaging quite a lot the next competition to learn about image segmentation is carvana image masking challenge very fun competition very nice competition and here you have like images of car and you have to segment it from the background so again i made segmentation problem and you can apply same kind of techniques that you have learnt in the previous one so in this competition about tgs salt identification challenge so you can apply the same techniques here um next we come to a challenge which is a combination of image segmentation object detection so you can you can use both of these in in this challenge which is the eye materialist fashion images challenge and this is a recurring challenge it happens every year so here i'm showing you the 2019 version but i think there was a 29 20 20 version too so your challenge is to given a fashion image you have to segment it in different parts different clothing or different accessories these kind of things and it's a it's a very nice challenge you're also given bounding boxes so you can also try to predict the bounding boxes so i would rather start from predicting the bounding boxes using some models like frcnn so here you will learn about again about units rcnns and all these kind of models for segmentation and here you don't have two classes you have many classes so that's what makes this challenge more interesting and these are images of people so um you the pro this is a little bit problematic because there are different poses and different uh types of images so but here you can learn about again about segmentation and about object detection and definitely you should take a look at this problem it's quite quite good quite good to learn about image segmentation image object detection then we have the global v detection nomination which is quite recent and here you have to detect the wheat heads so let me see if i have a notebook here where maybe maybe i have another book so maybe i can show you maybe you can learn how to use yolo uh in this one and even in the previous combination you can learn how to use yellow so this is what you have to do in the end you have to detect these bounding boxes so like fast rcn and base models can be used yellow can be used or any kind of other object detection algorithms can be used there have been a lot of new ones recently which are quite state of the art and this is what you will learn in this competition about detecting these wheat heads or object detection and when you're done with all these different kinds of competitions what you can do is you can also dive a little bit into gans and everyone is doing it so this is a good competition to start from and here you have to generate dog images um so basically you learn about all about cans different kinds of cans so um this is my list and i think this is a good list for beginners i have one dog images two flower classification three c discount for pneumonia five facial key points six right whale seven uh tgs eight carvana nine fashion uh ten uh wheat um head detection and the last one generative dog images for gans and i think this is a good list if you're a beginner and if you want to start from and you can start in this order and uh i hope you like this video if you do like do click on the like button and do subscribe and share it with your friends and i will see you next time in the advanced version of computer vision and also i'm making a list of tabular data composition and time series competitions so see you next time goodbye

Channel: Abhishek Thakur

Views: 10,120

Rating: 4.9814243 out of 5

Keywords: machine learning, deep learning, artificial intelligence, kaggle, abhishek thakur, best computer vision problems, best computer vision competitions, kaggle computer vision, computer vision beginners, beginners start computer vision, dogs cats kaggle, where to start computer vision, best computer vision competitions on kaggle, list of kaggle competitions, best kaggle competitions for beginners, best kaggle competitions

Id: 1-myowrUhok

Length: 17min 7sec (1027 seconds)

Published: Thu Dec 17 2020