Is instance segmentation support planned?

[Robotatron]

Is instance segmentation support planned?
YOLO5 now supports segmentation

[alexheat]

Hi @Robotatron I want to look into this. Would you be able to help? Can you share any resources on the Yolo5 segmentation with segmentation samples in other formats like Coco that would make sense to transform into Yolo?

[mosolab]

Hi! Thanks for tool.
I’m interested too and could provide you examples.
Please let me know how to share it.

[alexheat]

Hi thank you. Can you share links on this issue thread?

…

On Sun, Jan 1, 2023, 3:11 AM Oleksii Mosolab ***@***.***> wrote: Hi! Thanks for tool. I’m interested too and could provide you examples. Please let me know how to share it. — Reply to this email directly, view it on GitHub <#65 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC5OL7ZJS77LZSODA2WHPJ3WQFQ65ANCNFSM6AAAAAARDXXO7U> . You are receiving this because you commented.Message ID: ***@***.***>

[Robotatron]

@alexheat

Hi @Robotatron I want to look into this. Would you be able to help? Can you share any resources on the Yolo5 segmentation with segmentation samples in other formats like Coco that would make sense to transform into Yolo?

Sure. Here is an official YOLO5 repo to convert COCO segmentation to YOLO5 format (works with YOLO7 as well) https://github.com/ultralytics/JSON2YOLO/blob/master/general_json2yolo.py

If you'd like to test training a YOLO5 for segmentation it's exactly the same as for object detection but using train.py from their "segment" folder from the repo: https://github.com/ultralytics/yolov5/tree/master/segment

Let me know if there is other information I could help with

[alexheat]

@Robotatron , @mosolab , I have written a proposal for how I will implement the conversion of segmentation conversion. If you have time please take a look and let me know if you have any feedback #72

[alexheat]

I have release version .44 with support for converting Coco segmentation annotations to Yolo segmentation format. You can see a demonstration in this notebook https://github.com/pylabel-project/samples/blob/main/coco2yolosegmentation.ipynb

[sudheer-palaparambil]

Thanks @alexheat for your immediate response and time. I downloaded the latest pylabel 0.1.44 and tried with some Coco segmentation annotations (done in makesense.ai). I got some broken annotations in some images. Attached both actual image and PyLabel generated image.

[alexheat]

Thank you @sudheer-palaparambil , can you explain what you mean by broken images? I do not see what is broken. Do you mean the dotted lines? I don't see how those could have been added by PyLabel because it is only touching the annotations.

Can you provide the steps that you are doing to get those lines and the dataset like you did last time? Also, what tool are you using the rendering the images?

[sudheer-palaparambil]

Hi @alexheat,

I mean that gap in the bottom annotation line.
I executed the following source.

<---------- Beginning
from pylabel import importer
import os

os.makedirs("data", exist_ok=True)
#Specify path to the coco.json file
path_to_annotations = "data/project-22124.json"
#Specify the path to the images (if they are in a different folder than the annotations)
path_to_images = "images/"

#Import the dataset into the pylable schema
dataset = importer.ImportCoco(path_to_annotations, path_to_images=path_to_images, name="Segmentation")
dataset.df.head(5)

dataset.path_to_annotations = "data/yolo"
dataset.export.ExportToYoloV5(segmentation=True)[1]
<----------- End

I copied the labels generated from training\labels folder and uploaded it along with the images to roboflow. The above images are rendered in roboflow. For annotation I used makesense.ai

The first images in the previous mail is rendred when I uploaded images and Coco-JSON generated by makesense.ai

Thank you

[alexheat]

I see it now thank you. Can you share the original annotations as well? I want to investigate what is causing the gap by looking at the actual annotation coordinates

[sudheer-palaparambil]

Herewith attaching the dataset containing original annotations, PyLabel generated annotations and rendered images as well.
Dataset.zip

[mosolab]

@alexheat also on this now. I could convert it but would like also to visualise labelling.
Could converted labelling be visualised with pylabel?

[alexheat]

Thank you @sudheer-palaparambil , I have released v45 that fixes the issue. The issue was an extra space at the end of each line. I am not sure it it really broke the annotations or just the visual on roboflow but the issue is fixed. Thank you for your help to test it.

[alexheat]

@mosolab, I need to find another package that can visualize these kinds of annotations. But in the meantime you can use the site that sudheer recomended, https://app.roboflow.com/. You can upload the annotations and images and you will see a visualization for each image like this.

[mosolab]

@alexheat here is my example. looks like not full connected

[alexheat]

Hi, are you use using the latest version of the package: pip install pylabel==0.1.45

…

On Wed, Jan 18, 2023 at 11:42 AM Oleksii Mosolab ***@***.***> wrote: @alexheat <https://github.com/alexheat> here is my example. looks like not full connected [image: Screenshot 2023-01-18 at 20 41 04] <https://user-images.githubusercontent.com/4863793/213278758-74a9b23a-c90b-4dd1-8e0e-1e42897b39b4.png> — Reply to this email directly, view it on GitHub <#65 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC5OL7YHXSQLSGA6JGXD7U3WTBBT5ANCNFSM6AAAAAARDXXO7U> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[mosolab]

@alexheat, yes, you are right. I had 0.1.44.
Just checked 0.1.45 and it works and looks great!
Thanks for support!

[sudheer-palaparambil]

@alexheat, sorry for the delayed response. It is working pefectly now. Thanks, keep the good work.

[alexheat]

Thank you everyone

[Robotatron]

you are awesome, thanks!