When first tackling a computer vision problem, any engineer should first asks: should I implement a solution from scratch, from a paper or known method, or use an existing solution and fit it to my needs?
This question goes hand-in-hand with the offering of implementations in OpenCV. Luckily, OpenCV has very wide and extensive coverage of both canonical and specific computer vision tasks. On the other hand, not all OpenCV implementations are easily applied to a given problem. For example, while OpenCV offers some object recognition and classification capabilities, it is by far inferior to the state-of-the-art computer vision one would see in conferences and the literature. Over the last few years, and certainly in OpenCV v4.0, there's an effort to easily integrate deep convolutional neural networks with OpenCV APIs (through the core dnn module) so...
