The role of confidence threshold

[majar5c]

When applying Sahi on ultralytics model I experimented with different confidence thresholds. I noticed the confidence threshold defined in the Sahi model not only filters out low-confidence predictions but also enlarges the bounding box of some predictions.

detection_model = AutoDetectionModel.from_pretrained(
            model_type="ultralytics",
            model_path=MODEL_PATH,
            confidence_threshold=CONF,
            device="cuda" if torch.cuda.is_available() else "cpu"
)

I have attached a Jupyter notebook inference_for_ultralytics.ipynb that could reproduce the observation.

When a Sahi model is set to a confidence threshold of 0.35, these are predictions on an image they looks alright:

Then the same model with a confidence threshold of 0.01 would produce the predictions below on the same image. As expected it would produce a lot of low-confidence bounding boxes:

However, if looking closely on the second image, there are two bounding boxes that don't actually fit the car anymore, with some extra space on its left:

So confidence threshold even changes the size of the high-confidence bounding boxes. What is happening?! This does not quite align with my expectation on confidence threshold which simply filters prediction. If anyone know why this happens, please let me know! Thanks in advance