Repeat detections with InferenceSlicer

[robmarkcole]

Search before asking

• I have searched the Supervision issues and found no similar bug report.

Bug

The detections show a clear repeat pattern:

There are 720 detections, but when I check the confidence values, there are only 6 unique confidence values - i.e. appears detections are somehow mirrored to other locations on the image

Environment

• sv = 0.25.1

Minimal Reproducible Example

import os
import cv2
import numpy as np
import pandas as pd
from glob import glob
from dotenv import load_dotenv
import supervision as sv
from inference import get_model

# Load environment variables (e.g., from .env file)
load_dotenv()
api_key = os.getenv("ROBOFLOW_API_KEY")

# Load the Roboflow model
model = get_model(model_id="your-model-id", api_key=api_key)  # anonymized model ID

# Constants
PATCH_SIZE = 640
image_dir = "your-image-dir"  # replace with actual directory path
images = glob(f"{image_dir}/*.png")

# Read an image
image_file = images[1]  # change index as needed
image = cv2.imread(image_file)

# Define inference callback
def callback(image_slice: np.ndarray) -> sv.Detections:
    result = model.infer(image)[0]
    return sv.Detections.from_inference(result)

# Set up slicer
slicer = sv.InferenceSlicer(
    callback=callback,
    overlap_filter=sv.OverlapFilter.NON_MAX_SUPPRESSION,
    slice_wh=(PATCH_SIZE, PATCH_SIZE),
    thread_workers=1,
)

# Run inference with slicing
detections = slicer(image)
print(f"Number of detections: {len(detections)}")

# Annotate results
bounding_box_annotator = sv.BoxAnnotator(
    color=sv.ColorPalette.DEFAULT.colors[8],
    thickness=2
)
label_annotator = sv.LabelAnnotator()

labels = [
    f"{confidence:.2f}"
    for class_id, confidence in zip(detections.class_id, detections.confidence)
]

annotated_image = bounding_box_annotator.annotate(scene=image, detections=detections)
annotated_image = label_annotator.annotate(
    annotated_image, detections=detections, labels=labels
)

# Save results to CSV
csv_sink = sv.CSVSink("out.csv")
with csv_sink as sink:
    sink.append(detections, {})

# Load CSV into DataFrame for inspection
df = pd.read_csv("out.csv")
print(f"Detections saved: {len(df)}")
print(df.sort_values(by="confidence", ascending=False).head())

Additional

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!

[SkalskiP]

Hi @robmarkcole 👋🏻 have you tried to manipulate iou_threshold value?