334 - Training custom instance segmentation model using YOLO v8

This video walks you through the process of training a custom YOLO v8 model using your own data.
Code generated in this video is available here: github.com/bnsreenu/python_fo...
In this exercise, I use a public dataset that shows multiple classes for segmentation. This is the same dataset from tutorial 330 (Detectron2) - • 330 - Fine tuning Dete...
Dataset from: leapmanlab.github.io/dense-cell/
Direct link to the dataset: www.dropbox.com/s/68yclbraqq1...
Data courtesy of: Guay, M.D., Emam, Z.A.S., Anderson, A.B. et al. ​Dense cellular segmentation for EM using 2D-3D neural network ensembles. Sci Rep 11, 2561 (2021). ​
To prepare this dataset for YOLO, the binary masks were converted to the YOLO format. Please follow this tutorial to learn about this process.
( • 332 - All about image ... )
If you already have annotations in COCO format JSON file, for example by annotating using makesense (www.makesense.ai/) then the annotations can be imported to Roboflow for conversion to YOLO format. Otherwise, if you are starting from scratch, just annotate datasets on Roboflow. (roboflow.com/). You just need to upload your images along with the JSON file and Roboflow will convert them to any other format, in our case YOLO v8.
For information about YOLO models:
docs.ultralytics.com/models/y...

[@baraspilkova7986]

Thank you so much for your tutorials! Very informative and helpful

[@tamerzohdy8589]

Amazing job. Thanks a lot for this tutorial

[@chusamaayub5967]

Thank you for the informatic tutorial

[@gerokatseros]

Your videos are super helpfull!!! don't change anything!

[@DigitalSreeni]

Thank you!

[@vtjustin333]

Hallo Sreeni, I just love all your videos. And really thankful for you for putting all the efforts and for making us understand it in such a simlple way. Now I have a kind of unique case to find some kind of failures from images. So it would be great if I can connect with you and able to share some of the images with you. I am pretty sure that you can help me out on this.

[@salvadornunez23]

genial gracias x tu tutoriales saludo desde argentina.

[@WasiurRahaman-ex7yv]

sir your video is very great , i love it

[@ErtizaAbbas]

Good to see again . I will be very helpful if you make a tutorial on classification, segmentation of bio products, any can be content in honey, or any other natural liquid product. Thank you for your efforts 👌

[@user-fw4xz4ew1e]

thank youo for the great tut :] i have a question after the model successfully resized the images to 800x800 and generated corresponding masks of the same size, I'm curious about the possibility of upscaling both the resized images and masks to their original dimensions. Additionally, I'm interested in maintaining the accuracy of the predicted polygons during this upscaling process. Is there a feasible method for achieving this, and if so, how can it be implemented effectively?

[@AbdullahKhan-uw8mo]

Thanks Alot Sir

[@loganw621]

Hi, I have a orthophoto from a drone after it has stitched many individual images together. How would I then apply segmentation to this large TIF file? Want to eventually classify each tree in the image as healthy or diseased.

[@bijaynayak6473]

excellent

[@user-ih1ty2yr4j]

Thank you for your tutorial! It would be very helpful if you could provide tutorial on nucleoli segmentation on H&E images.

[@ghillesainouche3329]

Thank you for your video, quick question, How do you make the bounding boxes not show? so you get only the segmentations

[@SL-nx3df]

Thank you so much,it is amazing

[@DigitalSreeni]

You're welcome 😊

[@user-fp2dt3qb8o]

Great work. Big fan. Can you make video on cancer image segmentation if possible?

[@Sifeddinemek]

thanks alot

[@viraljangale4546]

Hey, do u know where us is the train.json file which is required for the annotation, in order to create instance segmentation

[@hamidgholami2683]

Hi I hope you are ok First of all thank you so much for these amazing tutorials I want to know whether you upload part 72 in this awesome series or not? If you do that I would be so grateful

[@Cicatka.Michal]

Díky!

[@DigitalSreeni]

Thank you.

[@md.sadiulalamchyon4336]

Thanks!

[@DigitalSreeni]

Thank you very much.

[@user-fk2jl2js3n]

one feedback is that you put a timestamp or at the start of the video the content divided in which parts of the video you will teach them. And thank for the content, its very clear, and I much appreciate it. Also I have this question, I notice that in the detectron2 video, you fine tune the model with 5 images for training and 2 for test, I would think that this is too little, it's posible tu fine tune a model with so little images? is there a min image?

[@DigitalSreeni]

Thanks for the feedback. I wish I had a team to take care of curating content and making it more presentable. I'll have to find time or focus on user experience for future videos. I use limited training data in my videos as I want to teach the process and not necessarily run an entire project. It is not a good use of my time to annotate 100s of objects that are required for accurate results. For real projects, the number of images is not that important; the number of objects is more important. I start with about 50 objects per class and train a model. Based on the results, I will add more annotations for objects giving me lower probabilities.

[@user-jn7li6yj4b]

How should I select a specific slide from a tif file?

[@guillermovelazquez3211]

Hi, thank you so much for your content Do you have any recomendation for a network to detect objects?, i am currently using yolov8x models but i would like to sacrifice sped over accuracy, i dont mind if it is slow, i have looked for RCNN variants, but it will be great if you can give me some names from your experience. Thank you, be well!!

[@DigitalSreeni]

I prefer Detectron2 (essentially Mask-RCNN).

[@alissonpina5339]

Great job, thanks! Do u think detectron is much better than yolo to detect submillimetre-scale cracks on concrete?

[@DigitalSreeni]

I will always rely on Detectron 2 first before finding alternate tools. Detecting cracks can be done using object detection approaches (e.g., Detectron) but it assumes that you have enough training data. You may want to consider anomaly detection approaches where you look at reconstruction errors based on good images (without cracks) to identify any deviations from ideal images. FOr example, with this approach: kzbin.info/www/bejne/hmqxlYR9lNt8rNE

[@viraljangale4546]

Hey, do u know where us is the train.json file which is required for the annotation, in order to create instance segmentation

[@prodbyryshy]

sorry if this is something you already explained, but this model i assume doesnt actually take 3D images as in 3 dimensional tiffs (height, width, depth), when it says it does 3d segmentation thats specifically referring to 3D scenes not actual 3D images, correct?

[@geospatial90]

Dear sir, Actually I want to classify tree species from high resolution image (12.5 cm). First of all I tried to experiment for individual tree crown segmentation using YOLO-8 model but in my case I have found multiple polygon for each tree instead of a single polygon. Can you please advise me why it happened?

[@shubhammakwana3736]

Hi sir, If In video The object will break in more number than will it show in same color or id.

[@soukainatariki229]

I used the same way to fine tune my model but l saw that the value of box loss increase with each element , it started with 0.700 and termined with 0.977 ! what l have to do ?

[@kvnptl4400]

Hi, this is semantic segmentation and not instance segmentation. For instance segmentation, there should be unique ID for same object class.

[@raihanvalentino2175]

Hello! in your video you display a mask_array according to mask number. How do I display the entire mask_array?

[@mahmoudma3n935]

We want videos on SegFormer

[@sarderiftekharahmed5558]

How can I modify the architecture of YOLOv8???

[@amrirasyidi]

I think the tensorboard doesn't work because you still use the default name for run directory, i.e. 'runs'. However in your code, you specify the project and name while training the model, hence creating customized run directory name, and it should be feed to the tensorboard

[@DigitalSreeni]

Yes, I realized that later. Apparently I am not good at troubleshooting while recording the video :)

[@viraljangale4546]

Hey, do u know where us is the train.json file which is required for the annotation, in order to create instance segmentation

[@sarderiftekharahmed5558]

Please make a video on modified YOLOV8

[@oanbui9154]

The segmentation results are quite bad. cell is not fully segmented.

[@vigneshvicky6720]

Plz explain the architecture of yolov8

[@FirstNameLastName-fv4eu]

ha ha ha :) someone who has done it for prod ready model - knows the pain