SSD: Single Shot MultiBox Detector

Idea

Another real-time object detection model faster than YOLO. The core of SSD is predicting category scores and box offsets for a fixed set of default bounding boxes using small convolutional filters applied to feature maps

Method

• Evaluate a small set (e.g. $4$ in the figure above) of default boxes (just like the anchors in Faster R-CNN) of different aspect ratios at each location in several feature maps with different scales (e.g $8\times 8$ and $4\times 4$ in (b) and (c) ).
• For each default box, predict both the shape offsets and the confidences for all object categories ($c_1,c_2,\dots ,c_p$)

Architecture

• The model is bases on a standard classification network
• Convolutional feature layers are added to the end of the backbone network. These layers decrease in size progressively and allow predictions of detections at multiple scales.
• For each feature layer of size $m\times n$ with $p$ channels, the basic element for predicting parameters of a potential detection is a $3\times 3\times p$ small kernel that produces either a score for a category, or a shape offset relative to the default boxes coordinates. At each of the $m\times n$ locations where the kernel is applied, it produces an output value.