SSD: Single Shot MultiBox Detector

SSD detects objects in images using a single deep neural network, discretizing the output space of bounding boxes into a set of default boxes over different aspect ratios and scales per feature map location. At prediction time the network generates scores for the presence of each object category in each default box and produces adjustments to better match the object shape, and by combining predictions from multiple feature maps of different resolutions it naturally handles objects of various sizes.

Because SSD is simple relative to methods that require object proposals, completely eliminating proposal generation and the subsequent pixel or feature resampling stage, it is easy to train and straightforward to integrate into detection systems. On PASCAL VOC, MS COCO, and ILSVRC, SSD achieves 72.1% mAP on VOC2007 test at 58 FPS with 300x300 input, and 75.1% mAP with 500x500 input, outperforming a comparable Faster R-CNN model in speed while providing a unified training and inference framework that is also more accurate than other single-stage methods.