Relative weights for each category in post-processing

[ccomkhj]

Context

• During postprocess, when multiple bboxes are merged, the category (class) of the highest confidence is assigned.
• However, when the bboxes size vary (some are small, some are big), then big bboxes are prone to be wrongly classified
• Imagine, apple disease detector with category (healthy-apple, disease-apple, flower), because of the partial observation, the bbox could be classified as healthy.

Suggestion

• Introduce relative weight among categories

classes = ['healthy-apple', 'disease-apple', 'flower'] 
get_sliced_prediction(
    image_path,
    detection_model,
    slice_height=int(crop_size[0]),
    slice_width=int(crop_size[1]),
    postprocess_match_threshold=0.2,
    postprocess_type="GREEDYNMM",
    postprocess_match_metric = "IOU",
    postprocess_class_agnostic=True,
    postprocess_relative_weight=[1, 3, 1],
    perform_standard_pred=True,
    overlap_height_ratio=0.3,
    overlap_width_ratio=0.3,
)

If it's relevant feature, I can make PR.