add automodel structure for unified model loading

[fcakyon]

New AutoDetectionModel class makes it easier to load any yolov5, mmdet, detectron2 model!

from sahi import AutoDetectionModel

# load yolov5 models
detection_model = AutoDetectionModel.from_local(
    model_type='yolov5',
    model_path='yolov5s.pt',
    image_size=640,
)

# load mmdet models
detection_model = AutoDetectionModel.from_local(
    model_type='mmdet',
    model_path='model.pt',
    model_config='config.py',
    image_size=640,
)

# load detectron2 models
detection_model = AutoDetectionModel.from_local(
    model_type='detectron2',
    model_path='model.pt',
    model_config='config.yaml',
    image_size=640,
)

[cemilcengiz]

Just using if source_is_video is enough

[cemilcengiz]

I think this registry should be better as class variable inside AutoDetectionModel. Also, why don't save the classes directly instead of saving only their name?

[fcakyon]

yes better move it inside the class.

saved only the names to avoid adding a new module import once a new detector model support is added. this way we only need to update a list of string (instead of adding module import).

can switch to saving directly the classes if its a better practice.

[cemilcengiz]

Why don't you want to avoid imports? Are they too expensive (I doubt it) or some of them require installation of other packages? As for the better practice, this way one might miss some bugs e.g. incorrectly written class names would not be noticed until the interpreter tries to instantiate them during run time.