On thing to keep in mind is that DeepStack will scale the image to the required input size of the model, that being 256x256, 416x416 or 640x640 depending on whether you have selected low, medium or high settings in DeepStack. Using the sub-stream as input to DeepStack therefore makes sense as it already matches the required input size fairly closely, while using the main stream wastes significant resource to first decode the main stream and the downsize the image afterwards. After an object has been identified though, saving hi-res alerts is certainly appealing.