[BUG] SAHI for mmdet3.0 does not reproduce the actual mmdet3.0 predictions.

[janglinko-dac]

Hey, any from the below

detection_model = AutoDetectionModel.from_pretrained(
    model_type='mmdet',
    model_path=model_path,
    config_path=config_path,
    confidence_threshold=0.0,
    image_size=512,
    device="cuda:0"
)
result = get_prediction(image, detection_model)

or

detection_model = AutoDetectionModel.from_pretrained(
    model_type='mmdet',
    model_path=model_path,
    config_path=config_path,
    device="cuda:0"
)
result = get_prediction(image, detection_model)

or

detection_model = AutoDetectionModel.from_pretrained(
    model_type='mmdet',
    model_path=model_path,
    config_path=config_path,
    confidence_threshold=0.0,
    image_size=512,
    device="cuda:0"
)
result = get_prediction(read_image(image), detection_model)

or

detection_model = AutoDetectionModel.from_pretrained(
    model_type='mmdet',
    model_path=model_path,
    config_path=config_path,
    confidence_threshold=0.0,
    image_size=512,
    device="cuda:0"
)
result = get_prediction(cv2.imread(image), detection_model)

is not equal to

model = init_detector(config_path, model_path, device='cuda:0')
result = inference_detector(model, image)

why?