Sliced Prediction question: How to prevent detection in bigger area than slice

[Makarand-F4F]

Hello,
I am very new to Sahi and have just recently started using it with a trained detectron-2 model. I have specified the sliced predictions and the model as below. The saved model that I am specifying was trained on 200 x 200 pixel images and I am using the same slice_height and slice_width for prediction.

detection_model = AutoDetectionModel.from_pretrained(
    model_type='detectron2',
    model_path="path",
    config_path="another path",
    confidence_threshold=conf_thresh,
    # image_size=640, # this I have commented out
    device = 'cuda:0',) 

  result = get_sliced_prediction(
      image_path,
      detection_model,
      slice_height = 200,
      slice_width = 200,
      overlap_height_ratio = 0.00,
      overlap_width_ratio = 0.00,)

Now, when I do the predictions, I expect the model to perform the object detection on each of the slice and then combine the results together but sometimes, I see a prediction that spans across multiple slices and using a much bigger tile. For examples, see the images below. First image shows correct small rectangles detected in each 200 x 200 pixel slice.

In this example image, you can see my detections as expected

In this image, you can see a much larger bounding box around an object that has been detected across much larger area than 200 x 200. How do I prevent this?

[fcakyon]

@Makarand-F4F you can achieve it by setting perform_standard_pred=False:

sahi/sahi/predict.py

Line 132 in 3c8f593

perform_standard_pred: bool = True,