By contrast, ScaleNet outputs a scale distribution in logarithmic space rather than a regressed single scale ratio, which helps the network converge to a reliable model [14]. In addition, the image box embedding that approximates the visible surface overlap measures is used to estimate the scale...