Decreased mAP 50 Results with YOLOv8 Model using sahi Evaluation Method

[Rayen023]

Upon using the same finetuned YOLOv8 Model, I have encountered a significant decrease in the mAP 50 results when evaluating the model on the same evaluation dataset using two different methods: "ultralitycs evaluate" and "sahi coco evaluate" with the no_sliced_prediction parameter set to True. The mAP 50 score dropped from 0.7 to 0.49, and I am seeking to understand the reasons behind this reduction.

Configuration

Here is my current configuration:

predict(
        model_type = "yolov8"
        model_path = <path_to_model>
        model_device = <device_type>
        model_confidence_threshold = 0.25
        source = <source_image_dir>
        slice_height = 512
        slice_width = 512
        overlap_height_ratio = 0.2
        overlap_width_ratio = 0.2
        dataset_json_path = <dataset_json_path>
        no_standard_prediction = False
        no_sliced_prediction = True
        image_size = 1088
        postprocess_match_metric = 'IOU'
        postprocess_type = 'NMS'
)

sahi coco evaluate --dataset_json_path <dataset_json_path> --result_json_path "$file_path"