Finally, the working version of YoloV8 Instance Segmentation

[sweetlhare]

Redid the previous pull request. Conducted testing.

[sweetlhare]

Test slice/no slice with detection model on image without objects and on image with objects.

[sweetlhare]

Test slice/no slice with Instance Segmentation model on image without objects and on image with objects.

[sweetlhare]

The interface is identical to the detection model

`from sahi.sahi import AutoDetectionModel
from sahi.sahi.predict import get_prediction, get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained(
model_type='yolov8',
model_path='yolov8n-seg.pt',
confidence_threshold=0.5,
device='cuda:0' # if torch.cuda.is_available() else 'cpu'
)

pred = get_sliced_prediction('small-vehicles1.jpeg', detection_model,
slice_height=640, slice_width=640,
overlap_height_ratio=0.4, overlap_width_ratio=0.4)

pred.export_visuals('test_results/', file_name='sliced_small-vehicles1-seg.jpeg')`

[Aprilistic]

Thank you for updating!! I managed to use yolov8 seg model to sahi. However, you missed threshold_confidence. At line 72, 73

Original, Line 72, 73
prediction_result_ = [(result.boxes.data[result.boxes.data[:, 4] >= 0.5],
result.masks.data[result.boxes.data[:, 4] >= 0.5])
for result in prediction_result]

Modified
prediction_result_ = [(result.boxes.data[result.boxes.data[:, 4] >= self.confidence_threshold],
result.masks.data[result.boxes.data[:, 4] >= self.confidence_threshold])
for result in prediction_result]

This code worked well.

[sweetlhare]

Thanks a lot! I set a fixed trash hold for local tests, forgot to change it back

[Jareco]

Probably I did something wrong, but I still have the following error:

File "C:\Projects\levasoft\SolarProTool-ObjectDetector\app\handlers\yolohandler.py", line 33, in predict
    result = get_sliced_prediction(
  File "C:\Projects\levasoft\SolarProTool-ObjectDetector\app\sahi\predict.py", line 261, in get_sliced_prediction
    prediction_result = get_prediction(
  File "C:\Projects\levasoft\SolarProTool-ObjectDetector\app\sahi\predict.py", line 100, in get_prediction
    detection_model.convert_original_predictions(
  File "C:\Projects\levasoft\SolarProTool-ObjectDetector\app\sahi\models\base.py", line 168, in convert_original_predictions
    self._create_object_prediction_list_from_original_predictions(
  File "C:\Projects\levasoft\SolarProTool-ObjectDetector\app\sahi\models\yolov8.py", line 182, in _create_object_prediction_list_from_original_predictions
    object_prediction = ObjectPrediction(
  File "C:\Projects\levasoft\SolarProTool-ObjectDetector\app\sahi\prediction.py", line 75, in __init__
    super().__init__(
  File "C:\Projects\levasoft\SolarProTool-ObjectDetector\app\sahi\annotation.py", line 581, in __init__
    raise ValueError("Invalid boolean mask.")
ValueError: Invalid boolean mask.

I use the following code:

    detection_model = AutoDetectionModel.from_pretrained(
        model_type='yolov8',
        model_path=config.YOLOMODEL_PATH,
        confidence_threshold=0.3,
        device="cpu", # or 'cuda:0'
    )

    result = get_sliced_prediction(
        "C:\\GRoof_22037481-dc2c-43d4-aab6-238d5f254bd6.png",
        detection_model,
        slice_height = 256,
        slice_width = 256,
        overlap_height_ratio = 0.2,
        overlap_width_ratio = 0.2
    )

Here is the picture:
https://imgur.com/a/GtcGS2Y

I detect obstacles on the roof, but somehow I get an error with invalid boolean mask. I also had this error with previous pull request.

At the same time, if I use another parameters, like

      result = get_sliced_prediction(
            "C:\\GRoof_22037481-dc2c-43d4-aab6-238d5f254bd6.png",
            detection_model,
            slice_height = 640,
            slice_width = 640,
            overlap_height_ratio = 0.4,
            overlap_width_ratio = 0.4
        )

then it seems to work, but result.object_prediction_list[0].mask.bool_mask contains 'False' values instead of points.

[sweetlhare]

I checked on your image and with your parameters, everything works. To get some detection results, I added an object to the image

[Jareco]

If I change some parameters, I get some results, but at the same time, I get as "mask" sometimes "False" values instead of numbers. See screenshots below:

[sweetlhare]

I will check what the problem may be asap

[rusvagzur]

Is this implemented in the main branch? I can't see the yolov8 segmentation model there

[rezat96]

Hey @sweetlhare, is there any update on this?
thanks

[OscarPetersen1992]

Any update on this?

[FourierMourier]

Any updates? masks are correct and it works fine

[CyVision13]

Thanks a lot, can you update please? @sweetlhare

Guys if you have problem with new version of ultralytics, you can change
from ultralytics.yolo.engine.results import Masks to from ultralytics.engine.results import Masks

[jehanjen]

Thank you for updating!! I managed to use yolov8 seg model to sahi. However, you missed threshold_confidence. At line 72, 73

Original, Line 72, 73 prediction_result_ = [(result.boxes.data[result.boxes.data[:, 4] >= 0.5], result.masks.data[result.boxes.data[:, 4] >= 0.5]) for result in prediction_result]

Modified prediction_result_ = [(result.boxes.data[result.boxes.data[:, 4] >= self.confidence_threshold], result.masks.data[result.boxes.data[:, 4] >= self.confidence_threshold]) for result in prediction_result]

This code worked well.
could you please the segment code with sahi

[jehanjen]

Is this implemented in the main branch? I can't see the yolov8 segmentation model there

the same for me

[jehanjen]

I checked on your image and with your parameters, everything works. To get some detection results, I added an object to the image

could you please share the cml for segmntation with sahi ans the code please

[jehanjen]

The interface is identical to the detection model

`from sahi.sahi import AutoDetectionModel from sahi.sahi.predict import get_prediction, get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained( model_type='yolov8', model_path='yolov8n-seg.pt', confidence_threshold=0.5, device='cuda:0' # if torch.cuda.is_available() else 'cpu' )

pred = get_sliced_prediction('small-vehicles1.jpeg', detection_model, slice_height=640, slice_width=640, overlap_height_ratio=0.4, overlap_width_ratio=0.4)

pred.export_visuals('test_results/', file_name='sliced_small-vehicles1-seg.jpeg')`

which code you have been used for segmentation with sahi?

[libazhahei]

I also wrote similar code to yours, but I found that the memory usage was very very high..

[jehanjen]

Ouod you please write which code you have been using

…

On Sat, 28 Oct 2023 at 5:50 AM, Shilong Li ***@***.***> wrote: I also wrote similar code to yours, but I found that the memory usage was very very high.. — Reply to this email directly, view it on GitHub <#918 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BBVNZNT45TUKTIFW2R7XQS3YBR6IJAVCNFSM6AAAAAA2TABKICVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTOOBTGY4DMOJVGY> . You are receiving this because you commented.Message ID: ***@***.***>

[sweetlhare]

Hello everyone!
In this version, segmentation should work fine, but there is really a lot of memory consumption with a large number of masks. The problem comes from the original logic of SAHI's work. Now, within our development team, we have made a simplified and optimized version, as soon as we finalize it for SAHI, we will update the pull request.

[alexkutsan]

I am also verry interested in YOLOv8 segmentation with SAHI.

Tried version from

pip install git+https://github.com/sweetlhare/sahi-yolov8-instance-segmentation.git@main

It processes several images and then throw an exception.

Traceback (most recent call last):
  File "/home/alex/rnd/company/tracks_tools/./segment_compare.py", line 75, in <module>
    check_detector("yolov8_sahi", yolo8_sahi_detector,
  File "/home/alex/rnd/company/tracks_tools/./segment_compare.py", line 54, in check_detector
    result = detect_func(image)
  File "/home/alex/rnd/company/tracks_tools/adssai/tracks_extraction/detectors.py", line 73, in __call__
    results = get_prediction(image, self.model)
  File "/home/alex/rnd/company/tracks_tools/venv/lib/python3.10/site-packages/sahi/predict.py", line 100, in get_prediction
    detection_model.convert_original_predictions(
  File "/home/alex/rnd/company/tracks_tools/venv/lib/python3.10/site-packages/sahi/models/base.py", line 168, in convert_original_predictions
    self._create_object_prediction_list_from_original_predictions(
  File "/home/alex/rnd/company/tracks_tools/venv/lib/python3.10/site-packages/sahi/models/yolov8.py", line 182, in _create_object_prediction_list_from_original_predictions
    object_prediction = ObjectPrediction(
  File "/home/alex/rnd/company/tracks_tools/venv/lib/python3.10/site-packages/sahi/prediction.py", line 75, in __init__
    super().__init__(
  File "/home/alex/rnd/company/tracks_tools/venv/lib/python3.10/site-packages/sahi/annotation.py", line 581, in __init__
    raise ValueError("Invalid boolean mask.")
ValueError: Invalid boolean mask.

Let me know if you need more info or more context.
i attached the video file that I am processing.

coldwater_10_sec.mp4

[jehanjen]

Hello everyone! In this version, segmentation should work fine, but there is really a lot of memory consumption with a large number of masks. The problem comes from the original logic of SAHI's work. Now, within our development team, we have made a simplified and optimized version, as soon as we finalize it for SAHI, we will update the pull request.

sahi segmentation with coco format??? can i use just the image to segment with sahi?

[fcakyon]

Hello everyone! In this version, segmentation should work fine, but there is really a lot of memory consumption with a large number of masks. The problem comes from the original logic of SAHI's work. Now, within our development team, we have made a simplified and optimized version, as soon as we finalize it for SAHI, we will update the pull request.

Wow amazing to hear that!

Feel free to ping me once PR is ready so that I can merge it 👍

[fcakyon]

yolov8 demo notebook should be updated with a segmentation example, failing tests need to be handled

[AntonioRodriguezRuiz]

Hi! Is there any updates on this?

[fcakyon]

Hi! Is there any updates on this?

We need @sweetlhare's cooperation to make this PR ready (fixing tests and adding/updating a demo notebook)

[gfx73]

Hi!
Up to my understanding the feature implemented here uses segmentation model to perfrorm detection. Am I right? I tried this code snippet and pred value's dictionary has Nones for masks.

The interface is identical to the detection model

`from sahi.sahi import AutoDetectionModel from sahi.sahi.predict import get_prediction, get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained( model_type='yolov8', model_path='yolov8n-seg.pt', confidence_threshold=0.5, device='cuda:0' # if torch.cuda.is_available() else 'cpu' )

pred = get_sliced_prediction('small-vehicles1.jpeg', detection_model, slice_height=640, slice_width=640, overlap_height_ratio=0.4, overlap_width_ratio=0.4)

pred.export_visuals('test_results/', file_name='sliced_small-vehicles1-seg.jpeg')`

[dimka11]

How to fix ValueError: Invalid boolean mask?

[AndrewArchie]

I was wondering if there will be a similar update for using segmentation for YoloV5? Thanks a lot!

[ozayr]

Hi, how would I get the actual polygon of the segmented area, currently only the mask seems to be available. Is this something that is there, or is it not implemented ?

[FilipeMendesVMD]

Does anyone has a colab notebook for this solution? I am trying to replicate but no success... :(

[jehanjen]

Hi, how would I get the actual polygon of the segmented area, currently only the mask seems to be available. Is this something that is there, or is it not implemented ?

i also need SAHI with segmentation for polygon ?

[fmandel98]

I included the provided code for segmentation and it works partially. Unfortunately is the bounding box prediction worse than before. It seems it is wrongly shifted. Someone has the same issue?

[kyra-smith]

Hi! Up to my understanding the feature implemented here uses segmentation model to perfrorm detection. Am I right? I tried this code snippet and pred value's dictionary has Nones for masks.

The interface is identical to the detection model
from sahi.sahi import AutoDetectionModel from sahi.sahi.predict import get_prediction, get_sliced_prediction detection_model = AutoDetectionModel.from_pretrained( model_type='yolov8', model_path='yolov8n-seg.pt', confidence_threshold=0.5, device='cuda:0' # if torch.cuda.is_available() else 'cpu' ) pred = get_sliced_prediction('small-vehicles1.jpeg', detection_model, slice_height=640, slice_width=640, overlap_height_ratio=0.4, overlap_width_ratio=0.4) pred.export_visuals('test_results/', file_name='sliced_small-vehicles1-seg.jpeg')

I'm having the same issue. Has anyone else found a solution for this?