How to use Double Detection Filter in Supervision library?

Community Help
1

I’m trying to follow the documentation and examples cited at effectively picking the higher confidence predictions for double detections.

and Double Detection Filter - Supervision

Below is an example image in question and my clarifications are regarding the two types of overlaps (encircled in orange and green)

Orange circle: What is the approach to pick the prediction with the higher confidence in this case?

Green circles: What some ways to think about thresholds that let us ignore the smaller enclosed predictions/ boxes?

It feels like I am mixing up the thresholds and the detection strategies.
Would someone mind clarifying the way to think about approaching detections in the above two scenarios?

Screen Shot 2024-06-16 at 10.30.32
Screen Shot 2024-06-16 at 10.30.32752×690 32.1 KB

This is current implementation.

def callback(image_slice: np.ndarray) → sv.Detections:
_, buffer = cv2.imencode(‘.jpg’, image_slice)
image_bytes = buffer.tobytes()

response = requests.post(
    f"https://detect.roboflow.com/{MODEL_ID}",
    params={"api_key": ROBOFLOW_API_KEY,
            "confidence": 45,
            "overlap": 10 },
    files={"file": image_bytes}
)
response.raise_for_status() 

response_data = response.json()
return sv.Detections.from_inference(response_data).with_nms(threshold=0.1)

image = cv2.imread(image_path)
slicer = sv.InferenceSlicer(
callback=callback,
slice_wh=(1000, 752),
overlap_filter_strategy=sv.OverlapFilter.NON_MAX_MERGE
)
detections = slicer(image)

For full context:
4000x3008 image is sliced into 1000x752 tiles for inference

2

Hi @kedar :wave:

If there are cases of clear overlap such as the green circles, I’d start with those first. Filtering out misdetections & cleaning the data is a good first step.

Since you wish to merge with the larger box, non-max merge and a threshold of about 0.1 - 0.05 is the correct approach. I’d run over multiple examples of your data until I can narrow down on a realistic number.

If this does not work, I’d look into a custom solution. For example, it seems like your model produces boxes that are confined within others. You could create a system that compares each boxes by how much one is NOT encapsulated in another. Then given 2 detections, if you check that A is almost entirely in B, it’s likely to be the green circle case.

For the orange circle, I’m not certain of what you wish to detect. Can you explain a little bit more?

3

Sure.

  1. As a clarification, is using with_nms(threshold=(0.1-0.05) here your suggested approach? The reason I’m a bit confused is I’m trying to better understand how this works in conjunction with the overlap threshold specified as %.

  2. The orange circle is an example of two predictions for the same object with different confidence scores. Behind the black bounding box is a purple one occluded.

Many thanks for your guidance @linas :pray:

4

I’ve updated a blog post a few days ago, on detecting small objects. Let me know if that explains it well enough:

Then, there’s an in-depth article by Piotr, explaining non-max suppression step-by-step:

1 Like
5

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.