Deep Learning for Computer Vision

Single shot multi-box detector

SSD (Single shot multi-box) is proposed by is the fastest of all the methods. This method simultaneously predicts the object and finds the bounding box. During training, there might be a lot of negatives and hence hard-negative mining the class imbalance. The output from CNN has various sizes of features. These are passed to a 3x3 convolutional filter to predict bounding box.

This step predicts the object and bounding box:

Reproduced with permission from Liu et al.

These are the algorithms available for object detection and we will learn how to implement them in the following section.