Full inference (without slicing) gives different results than YOLOv8

[andressrodrl]

I am using the same weights, confidence threshold, IOU treshold and size, also choosing 'NMS' as post-process method in SAHI. But the results are different and the detected objects are lower on SAHI. Any clue?

EDIT: looking it closely, the predictions are actually the same, the problem is that SAHI doesnt actually uses the model_confidence_threshold im defining, so it doesnt detect as much. It seems it uses 0.25 as threshold even tho i set it to 0.01

[wuyifan2233]

I am using the same weights, confidence threshold, IOU treshold and size, also choosing 'NMS' as post-process method in SAHI. But the results are different and the detected objects are lower on SAHI. Any clue?

EDIT: looking it closely, the predictions are actually the same, the problem is that SAHI doesnt actually uses the model_confidence_threshold im defining, so it doesnt detect as much. It seems it uses 0.25 as threshold even tho i set it to 0.01

Hi, have you solved this problem? I face the same situation with you.

[GCGeo]

It appears certain configuration parameters for the detector are only applied when sliced prediction is on (no_sliced_prediction = False).
I had a similar problem with changing the confidence threshold, and also when trying to set class-agnostic NMS to True. The parameters are not passed on to the detector.
I have tried passing the parameters directly to the batch predict() function, passing a config.yaml file with the parameters in the predict() function, and also creating a AutoDetectionModel with the set parameters and then passing the detection model to the predict() function, none of that worked.

[WaderLaken]

I also get an incomprehensible problem using SAHI with YOLOv8 model.

When I compute the metrics for val datesets using standard inference with a YOLOv8 Model, I could get 0.862 mAP50. However, I just get 0.607 mAP50 using sliced inference with the same model, simultaneously setting "no_sliced_prediction=True" or setting "no_sliced_prediction=False" but tuning the "slice_size" to an oversized value, like 8000x8000, definitively inferring only one slice on a picture. By the way, I ensure the other reasoning variables the same. What a surprising result!

What's more, I make another try, using sliced inference with am appropriate slice-size (such as 4032x4032), and the mAP50 rises to 0.931. Now I confirm that SAHI is effective to improve the recognition accuracy, especially for tiny objects. However, I really wonder why the values between the last two experiments are markedly different?

[fcakyon]

there was a bug in YOLOv8 prediction thresholds. This PR fixed the issue: #988

@andressrodrl @wuyifan2233 @WaderLaken @GCGeo