latest/detection/core/

[giscus[bot]]

latest/detection/core/

A set of easy-to-use utilities that will come in handy in any computer vision project.

https://supervision.roboflow.com/latest/detection/core/

[jeffrymahbuubi]

Hi, I was wondering if the sv.Detections.from_ultralytics function would work with the stream=True argument, as the results will return a generator. I tried using sv.Detections.from_ultralytics, but it seems it can't process this generator. I wonder if there is any added benefit to using stream=True when combining Supervision with the Roboflow Inference library.

def predict(self, image: np.ndarray) -> sv.Detections:
    results = self.model(image, verbose=False, device=self.device, stream=True)[0]
    return sv.Detections.from_ultralytics(results)

[LinasKo]

Hi @jeffrymahbuubi 👋

My understanding is that stream mode is typically used for videos and streams to avoid loading them into memory. We don't support stream in Ultralytics, but have a few alternatives.

For videos, you can use get_video_frames_generator. Our typical recommendation for streams is to use InferencePipeline with Roboflow models instead.

[olawalejuwonm]

This is a great library. I've used it for a research project, but I'm curious to know if there is any paper on it that i can cite?

[LinasKo]

Hi @olawalejuwonm 👋,

Glad to hear it's useful! If you go to the Repository page, on the right-hand side, you'll see a "Cite this repository" button. That should provide a copyable snippet for you.

Thanks again for the kind words!

[fakurten94]

Is there a way to get the mask XY coordinates from the output Detection object? As it outlines, you can only retrieve the XY coordinates of the bounding box but when using a segmentation model I wasn't quite sure if you could. If it can't it would be a really interesting feature to add as the YOLO models you can retrieve those coordinates.

[LinasKo]

Hi @fakurten94 👋

Check out sv.mask_to_polygons. It converts each detected mask to a set of xy points. Is that what you were looking for?

[MrMerowinger]

Hello roboflow,
I have an issue when adding new key-value pairs to the data dictionary in sv.Detections.

for image_path, image, annotations in detection_dataset_train:
    for xyxy, mask, confidence, class_id, tracker_id, data in annotations:
        data["category"] = ["dog"]
        data.__setitem__("position", [POSITION.CENTER])
        print(data)     

{'category': ['dog'], 'position': [<POSITION.CENTER: 'center'>]}

Its really odd that when I want to access it, the added values disapear.

for image_path, image, annotations in detection_dataset_train:
    print(annotations.data)

{ 'category': [0], 'position': [0]}

What is reason for this and how can I add new pairs to the Detections after initialization? Thanks in advance!

I fixed the issue by saving the updated 'data' back to the annotations

annotations.data = data

[SkalskiP]

Hi, @MrMerowinger 👋🏻 Set your categories like this:

annotations["category"] = ["dog", "cat", ...]  # The list should have the same number of elements as len(detections)