Automation and Classification Cascade Methods for Object Detection : Organising Data Stack for Supervised Learning before Runnning Unnecessary Pathfinders
Pedro Faria, Kai Miao, Tomás Simões, Qianmin Yan · 2024
Let's assume this scenario, that a machine learning project has grouped 300 classes for each object type to detection, running common algorithms such as a Random walk in which every class can actually be bind to each other in small or larger clusters. A total possible combination for this use case would be around 2.037×10^89 possibilities of different groups. It’s not the amount of combinations that is in question here, but the time and energy consuming to compute this data, every time there would be a new object unsupervised to be properly classified. Therefore, the purpose of this paper aims to help machine learning beginners or enthusiasts, on their process of data stack structure, precisely at the beginning of the whole process, using as a case study of inside housing detection that uses more than 300 classes. Some examples and methodologies, including calculation models, are also cited with the final purpose to clear planning and guidance. The reason substantiated is the fact that many new machine learners, mostly students, are simply eager to run their code to get their object detection classification, directly to the purpose they want to achieve, no matter the path the machine is obliged to go until it shows some results. This anxiety is not only part of youth but part of the self-learning process when the machine starts to declare either is taking too long or the cost is becoming unbearable. Nevertheless, it’s a pain that can be avoided and the stack organisation helps developers additionally on defining better their patterns and create an automation process so that the future data will also just self-organize easily. There will be guidance using some previous works with some examples and algorithms to fully understand how simple, at the beginning, it is to organize data, and to better teach our machines, before just launching and learning at all cost. It is obvious that machines don’t complain and eventually they learn, however the following paper, shows graphically the method in an abstract way, as well as an example of inside house detection, hoping to be a great contribution to save time and to organise new projects for developers, at foundation level.