YOLOv5 segmentation now is possible! #46
Conversation
There was a problem hiding this comment.
Thank you for your contribution.
I think you have skipped all pre-commit when committing your code. Please do following:
- Please run
black -l 79 .to format your code. - Run
pre-commit run --all-filesand refactor your code to follow this repo's standards. - Use uppercase character for the first character of all comments.
- Use models here: https://github.com/vietanhdev/anylabeling-assets/releases/tag/v0.3.0.
| if area > avg_area * 0.2 | ||
| ] | ||
| contours = filtered_approx_contours | ||
|
|
There was a problem hiding this comment.
Please add the first point again as the last element to close the contour.
Like the line 271 of this file: https://github.com/vietanhdev/anylabeling/blob/master/anylabeling/services/auto_labeling/segment_anything.py
points.append(points[0])
The contours will not be closed if you don't do it.
There was a problem hiding this comment.
|
There are some differences between your YOLOv5 segmentation result with the result from this repo (using the same model). For YOLOv5n Segmentation
Note: Missing chairs and not good segmentation result in the man's leg. For YOLOv5x Segmentation
Could you please check the implementation for the reason? |
|
Thanks @vietanhdev, Please, could you share the image you are using? so both, can work on the same source. |
|
Please check this image. |
In this PR you will be able to use YOLOv5 segmentation from Ultralytics to create masks and accelerate the process in labeling data.
All the segmentation models can be downloaded from here: https://github.com/hdnh2006/anylabeling-assets/releases/tag/v0.0.0
They have been exported with
--opset 12as option as following:python export.py --weights yolov5x-seg.pt --include onnx --opset 12An important change has been done in the
blobpreprocessing of all yolo models, since the image is already in BGR format, soswapRB=Falsehas been set in order to avoid changes in the channels.Again, if there is any change you need, please let me know.