205 - U-Net plus watershed for instance segmentation

Поделиться
HTML-код
  • Опубликовано: 17 ноя 2024

Комментарии • 73

  • @durarara911
    @durarara911 Месяц назад

    Dude, you're great! You explain everything in detail, and I love how you visualize the steps, and are patient enough to go through them all. I've learnt so much from you!

  • @mostafanakhaei2487
    @mostafanakhaei2487 8 месяцев назад

    This is a great video for explaining traditional computer vision, keep up the great work

  • @liamrussell2192
    @liamrussell2192 3 года назад +1

    Thank you so much for all your content, as a fairly new microscopist your whole channel has been so incredibly helpful!

    • @DigitalSreeni
      @DigitalSreeni  3 года назад +1

      Welcome to the wonderful world of microscopy and image processing.

  • @tonihullzer1611
    @tonihullzer1611 2 года назад +1

    Wow, so helpful. This is the only place where you can find so special knowledge so perfectly presented. Thx a lot.

  • @taarieq2727
    @taarieq2727 3 года назад

    Thank you for the video. I have been trying to wrap my head around how I can convert my predicted masks to instances for the calculation of the panoptic quality of my model. This was the missing key. Thanks!

  • @onurozel4310
    @onurozel4310 3 года назад

    This is my first comment in RUclips. Mr. Sreeni, your videos is perfect. Your information is so clear, and your pronunciation is so good. I learn many thing about AI - speacially image segmentation with u-Net- through by your videos. Thanks a lot Mr. Sreeni . You are the best Teacher I have ever seen.

  • @asmabenbrahem6
    @asmabenbrahem6 3 года назад

    Thank you very much for these amazing tutorials, very well explained. Keep up the good work. May God bless you.

  • @RobertReihs
    @RobertReihs 3 года назад

    Thank you for this great video. It is a pleasure to learn from your tutorials. And amazing how much effort you put in creating this videos.

  • @spyponders275
    @spyponders275 3 года назад +5

    Thank you for the effort you put into these videos, I have learned so much from you!
    One question if you don't mind. How hard is it to expand the watershed here to 3D images (confocal volumes), is it possible, what would have to change? I think I understand how watershed works on 2D images but I have trouble visualizing how it would work in 3 dimensions. Thanks again!

  • @tonihullzer1611
    @tonihullzer1611 2 года назад

    i noticed after some time doing computer vision it is really necessary to get to know in depth the cv2 library and the traditional image processing!

    • @DigitalSreeni
      @DigitalSreeni  2 года назад

      Learning cv2 is a prerequisite to learning any machine learning, whether it is scikit learn or tensorflow.

  • @milanpavlovic4415
    @milanpavlovic4415 3 года назад +3

    Your videos are very inspirational, thank you!
    What would be your approach if you want to do instance segmentation to differentiate close objects, but in some cases, they can overlap?
    Is there something like multiclass instance segmentation? I'd love to see your video on that!

  • @xiaoli4056
    @xiaoli4056 3 года назад

    another pure gold video, Thanks !

  • @chuankangpei7279
    @chuankangpei7279 2 года назад +2

    Thanks for your wonderful video! It really helped a lot!
    I have one question about evaluating the performance of the U-Net plus watershed for instance segmentation - I know that I can evaluate UNet results using some metrics like Mean IoU, do you have any suggestions on how do to quantitatively evaluate the performance of unet marked watershed instance segmentation results (that the size of an object is properly determined)?
    Thank you very much!

  • @alexandergrishin7084
    @alexandergrishin7084 3 года назад +1

    Hi! Very nice explanation, thank you! I have a question. What if instead of setting some threshold for the model's output and then defining your areas of confirmed background and foreground by processing the image, use different model's thresholds for obtaining background and foreground, to begin with?

  • @oleg.bissing
    @oleg.bissing 10 месяцев назад

    Thank you for this amazing tutorial

  • @venkatesanr9455
    @venkatesanr9455 3 года назад +1

    Hi Sreeni Sir, Always doing great for the community through your videos. Thanks for your videos and I have learnt a lot from your informative content. I have only 250 images for my segmentation experiment whether I can try U-net or go with ML approaches like Random forest/Xgboost. Kindly suggest your experience

  • @MinhNguyen-yz3fy
    @MinhNguyen-yz3fy 3 года назад

    Thank you very much for sharing knowledge. I learned a lot from your videos. I especially like your videos about U-Net. Can u make a video about generating RGB photos from RAW images using the U-Net.

  • @carlosleandrosilvadospraze4005
    @carlosleandrosilvadospraze4005 3 года назад

    Great video! Thanks for sharing your knowledge!

  • @adeelazam5762
    @adeelazam5762 3 года назад

    You doing a great job on image processing and machine learning but i need some help regarding video processing. if you make a short video it would be beneficial for all who are interesting in this filed.

  • @willberger96
    @willberger96 Год назад

    What a great video! Thank you!

  • @vladlt6846
    @vladlt6846 2 года назад +1

    Great video.
    I'm interested, do you have any proof or discussion for the statement you mentioned earlier that the Unet is better than MaskRCNN when the dataset is more scientific than common?

  • @amnesie148
    @amnesie148 3 года назад

    very detailed explain, a great video thank you very much!

  • @XX-vu5jo
    @XX-vu5jo 3 года назад

    This is good if datasets are easy to get and are widely available

    • @DigitalSreeni
      @DigitalSreeni  3 года назад

      Isn’t that the case with every supervised deep learning? Bulk of the time is spent in labeling but once you have labeled data you are good to experiment with many techniques. There are many companies making a living by providing labeling services just because the task is very time consuming but the value of labeled data is incredible. You can try transfer learning by using a pretrained encoder in the Unet. I’ll do a video on it but you still need a good amount of labeled images from your own dataset.

    • @XX-vu5jo
      @XX-vu5jo 3 года назад

      @@DigitalSreeni please do video on on knowledge distillation or dark knowledge too!! :)

  • @dungeondragon9329
    @dungeondragon9329 Год назад

    Thank you for this great video

  • @aankitdas6566
    @aankitdas6566 3 года назад

    THIS IS IT! AMAZING CONTENT

  • @kannanv9304
    @kannanv9304 3 года назад

    Pranams Ajarn.......Have just started watching it........Meanwhile, I am working on 204.....Little bit stuck with "Chopping/Cropping" images, but I will get through that......

    • @DigitalSreeni
      @DigitalSreeni  3 года назад +2

      I will be talking about a library that helps with dividing images into patches, so please do not sweat too much trying to do things that can be easily automated :)

  • @giulsdrll
    @giulsdrll 3 года назад

    Interesting! Thank you for Your work

  • @zaheerbeg4810
    @zaheerbeg4810 3 года назад

    Thanks Sir , keep posting

  • @tonihullzer1611
    @tonihullzer1611 2 года назад

    should the boundaries that are decoded as -1 be treated s they do belong to the background or are they part of the cells?
    what i now do for instance segmenation is, i transfer all the markers = markers -10 back and all -1 to 0, so background is 0 again, and what remains are the pixel with the same value belonging to one instance.

    • @DigitalSreeni
      @DigitalSreeni  2 года назад

      I normally assign boundary pixels as part of the object and not the background. Considering they are only made up of single pixels, it does not matter whether they fall into object class or the background class.

  • @harrishvar7677
    @harrishvar7677 9 месяцев назад

    Hello sir, Will this approach apply if i want to segment farmlands from satellite images using u-net?

  • @mohammaddehghan8762
    @mohammaddehghan8762 2 года назад

    just i can say, very thank you

  • @rezatabrizi4390
    @rezatabrizi4390 3 года назад

    Hi, I'm very happy to follow your channel. I have a question about differences between Unet and other techniques such as Tiramisu layers or pyramid layers. Which are better for semantic segmantation?

    • @DigitalSreeni
      @DigitalSreeni  3 года назад +2

      There must be a reason why Tiramisu based segmentation is not around anymore. I believe it came out in 2017 or so but apparently it worked great on CamVid dataset and failed on real life biomedical and remote sensing type of images. As of today (early 2021) U-net is the king when it comes to semantic segmentation. It worked for me very well on various types of data so I am not even exploring other approaches unless I see a paper that proposes radically new approach. Same with instance segmentation, I do not like other approaches for most tasks as Unet with watershed does a good job for the many use cases. You will always find a nice area where something else works but for biomedical images this approached worked great for me.

  • @Lemonjjuice
    @Lemonjjuice 3 года назад +1

    For a large image splitted into patches, do you recommend to merge them back to the original image and apply the watershed algorithm or the other way around?

    • @DigitalSreeni
      @DigitalSreeni  3 года назад +4

      First merge into a large image and then watershed.

  • @talha_anwar
    @talha_anwar 3 года назад +1

    Can you please introduce image augmentation in next video of segmentation

    • @eli_m6556
      @eli_m6556 3 года назад

      He actually has a couple videos on augmentation, look back through the channel. Also, look Albumentations library, best library for augmentations.

    • @talha_anwar
      @talha_anwar 3 года назад +1

      @@eli_m6556 thanks

  • @adeelazam5762
    @adeelazam5762 3 года назад

    It is my request can you make one video on python related to small Video processing with following short agenda
    1. load input video and extract frames from video using some python libraries
    2. basic operation on video frames
    3. write video and save to directory
    I am waiting for your video. because i will start my work on video processing using python. Thank you very much. Am much awaited for you upcoming video.

    • @DigitalSreeni
      @DigitalSreeni  3 года назад

      Thanks for the suggestion. I will add it to my list. In fact, I tried to do it in the past but I was unable to find a good video file that I could use.

  • @sabeenak7159
    @sabeenak7159 3 года назад

    Hats off u sir🙏

  • @sachinchandrasekra4942
    @sachinchandrasekra4942 3 года назад

    I need to calculate the Brain tumour area. can you explain that how to do that?

  • @questless3033
    @questless3033 2 года назад

    how to do recurrent instance segmentation? can I use u-net ?

  • @МарияЕремина-я6г

    Thank you so much!

  • @shilpashree4860
    @shilpashree4860 3 года назад

    Great work sir, Can u make a video on multiclass image segmentation using UNET model. Is it required to use one hot encoding for multi class segmentation.

    • @DigitalSreeni
      @DigitalSreeni  3 года назад

      Multiclass Unet is coming soon. Yes, you need one hot encoding.

    • @hemingliao1917
      @hemingliao1917 3 года назад

      @@DigitalSreeni Am looking forward to this too. Thank you very much for your video.

    • @mimolinodeviento
      @mimolinodeviento 3 года назад

      @@DigitalSreeni looking forward to that video!! I hope it comes out soon :)

  • @timmyrex935
    @timmyrex935 3 года назад

    thank you for this video

  • @danicalifornia94403
    @danicalifornia94403 3 года назад

    Can we track the each mitochondrion (or cells) from live-imaging data with the instance segmentation by the code you made?

    • @DigitalSreeni
      @DigitalSreeni  3 года назад +1

      Yes. This video explains the process of segmenting each mitochondria (or cells) into individual objects. Now, you need tracking algorithms (e.g., Brownian model) to track each object. We created a workflow for exactly this on APEER. www.apeer.com/app/workflows/detail/Object-Tracking:-End-to-end-including-object-segmentation-and-separation-steps/2d29df90-a50c-42b8-9560-28ae5cb3094b

    • @danicalifornia94403
      @danicalifornia94403 3 года назад

      @@DigitalSreeni Thank you so much!

    • @danicalifornia94403
      @danicalifornia94403 3 года назад

      @@DigitalSreeni Can you make a video about what you explained?

  • @DesertWolf90
    @DesertWolf90 2 года назад

    could you please upload the script for multi images ?

  • @hichammaarouf6095
    @hichammaarouf6095 3 года назад +1

    hi sir new follower
    please i couldn't find the file where y applied this on dataset !!!

    • @eso7565
      @eso7565 2 года назад

      did you find the files? i need them too :)

  • @tonihullzer1611
    @tonihullzer1611 2 года назад

    Danke!

    • @DigitalSreeni
      @DigitalSreeni  2 года назад

      Thank you very much. Please keep watching...

  • @adelbenali4308
    @adelbenali4308 9 месяцев назад

    the masks supplied with the images how to prepare them please