Themaximumnumber of hardmining examples was set to 3,000,and the ratio between negative and positive examples was left atthe default value of 3:1. A positive example is a proposed boxwith an annotated object of interest and correct box scale. Anegative example is a proposed box with no annotated objectsof interest and an incorrect box scale. The default box generatorwas applied to 6 different convolution layers with a minimumand maximum scale of 0.2 and 0.95 respectively. The defaultboxes were generated with fixed aspect ratios 1.0, 2.0, 0.5, 3.0, and0.333. The complete details of the SSD model design principlesare provided in Liu et al. (2016). In order to perform real timeinference on a mobile device, images were resized to 300 × 300pixels before being fed into the network.