Multiclass Segmentation using UNET in TensorFlow (Keras)| Semantic Segmentation | Deep Learning

In this video, we are working on the multiclass segmentation using UNET architecture. For this task, we are going to use the Oxford IIIT Pet dataset, which consists of three classes:
1. Main object (Cat or Dog)
2. Border
3. Background

What is semantic segmentation?
The goal of semantic image segmentation is to label each pixel of an image with a corresponding class. It is also called Dense prediction.

What is U-Net?
U-Net is a fully convolutional neural network that was developed by Olaf Ronneberger. It was specially developed for the purpose of biomedical image segmentation.

Support:

MY GEARS:

FOLLOW ME ON:

[gamefever]

Thanks for showing all bugs and not cutting them from the video.

[priyankaarora]

Very useful for anyone working with multiclass segmentation.... keep it up @nikhiltomar.... god bless you

[nicopaolinelli]

Thank you very much for your videos! You are amazing! I shared my work based on semantic segmentation on kaggle and you've been very precious. I also mentioned you in the notebook!

[niluparupasinghe]

This is very helpful. Highly recommend this solution for anyone looking to create a data pipeline to fed in to a segmentation model training without exhausting resources like the RAM. Thank you very much for sharing, keep up the good work!....

[lawrencepratt]

This is great stuff, and right on point for my project.

[YTGmeathead]

Amazing video! Explain this topic in a such clear simply way! Thanks very much!

[husammasalkhi]

Dude amazing video!!
So compact in your coding and at the same time very general for any application. Some other tutorials use pre-built datasets or use a prebuilt encoder, but you do it from scratch, top

[aquienleimporta]

i had to do some changes for my adaptation but man, it works!!! Thanks a lot, you're a genius

[atharvahude]

Thank you very much for the video I wanted to do multiclass segmentation using Unet but was getting shape errors your code helped me !

[matthewavaylon]

Questions:
1. Why do we read the mask as float32?
2. Why do we subtract 1 from the mask when we read it
3. Can you explain why we use "decode" in the preprocess function.
4. I'm adapting your code with my CRF masked as an RNN which needs a batch size of 1. Do you think the accuracy will be affected if I change the batch size to 1 instead of 8?

[HungNguyen-ddiw]

Thank you for your helpful video! I haven't understood difference between 1 class and multiclass in code ( or in program). In this video, UNet had 3 classes. I saw that you filled '3' in some parameters. But in dataset, must we split dataset to 3 parts (or 3 folders: cats, dogs and something?) and shall we assign them to 3 ID? I learn a lot of tutorials but they almost train single class Unet. In some cases, people only changed number of class in code ( or in program, in function). No one told anything about the classes in dataset or difference between single class and multiclass. I'm very confuse because I think that if you only change parameters from '1' to 'n (class)' in code, it's so simple and nothing difference between single and multi class. I have a project using multiclass Unet but I can't understand it. Please, can you explain me!. Thank you very much!

[igeek]

hey why do we need to store file names in separate text files? Why dont we process directly with the images rather than collecting with filenames from texts?

[sohailali]

Wow great tutorial thank you so much.
Can you please make a tutorial on how to evaluate trained model?

[RethinkerMedia]

Will dice coefficient and soft dice loss work for the multiclass UNET segmentation?

[eshrattrisha]

I have a dataset with lung+infection mask. But it is binary masking. I do not have any csv or json file with it. Now, how should I apply unet? If I also want lung+infection mask using unet, should I do multiclass segmention?

[djdekabaruah]

Hi..great tutorial.. could you please make a video on Kaggle "Aerial Semantic Segmentation Drone Dataset"? It is for multiclass semantic segmentation.

[guilhermematos]

One question: to train networks by segmentation is it necessary to have ground truth images? And how can I produce ground truth images on my own with a dataset I created?

[vijayaprashanth]

Thanks for the informative video sir. I tried to add one custom loss function but unable to achieve as the ytrue shape is being taken as (none, none, none, none)

[guilhermematos]

One question, why is it necessary to do this division: (train_x, train_y), (valid_x, valid_y), (test_x, test_y)? Is it always necessary and for every dataset, or optional for my own dataset?

[manueladdae]

Please make a video on a transformer Unet hybrid.