Closed changchiyou closed 8 months ago
Most detectors can work with different input sizes, within a reasonable range. The choice of detector's input size (det_size
) is based on some common sense judgement:
@kormalev That makes sense. It seems like I need to strike a balance between speed and accuracy, right?
What does
det_size
actually do? I can't find any document talking about it. https://github.com/deepinsight/insightface/blob/01a34cd94f7b0f4a3f6c84ce4b988668ad7be329/python-package/insightface/app/face_analysis.py#L47I have do some research by myself. My understanding is that the photo will be scaled based on
det_size
, and this is the sole purpose ofdet_size
. But I am not sure whetherdet_size
is a fixed number for each detection model or not.https://github.com/deepinsight/insightface/blob/01a34cd94f7b0f4a3f6c84ce4b988668ad7be329/python-package/insightface/model_zoo/retinaface.py#L207-L221
BTW, there are acutally MANY version of
det_size
:This makes me feel confused, and I would be concerned that arbitrary changes to the default values might lead to a decrease in the model's accuracy in judgment.