Object detection vs. semantic segmentation on fashion RGB images

Abstract

[en] People have loved clothing fashion for thousands of years, from the early days of Egypt until nowadays. Throughout history, drawings, documents, and other archaeological finds have also revealed that people wore fashion in different moments of history. As an example, we have had various civilizations. The Greeks wore thick woolen long dresses. The ancient Egyptians were typically dressed in light cotton clothing. The Romans became the most critical example of style and fashion because of their expansion and dominance. In the recent past, the fashion industry has emerged as one of the crucial industries for the global economy. The trends change every second, the clothing industry has proved itself one of the most creative realms, and with the advent of the internet and handheld devices, customers can easily shop on the go. While people keep up with fashion trends, machine learning is changing the trends in the fashion industry, and daily there are systems keeping track of every sale and the upcoming trends. This gives the companies vast knowledge about what a user is interested in. Seeing how many opportunities there are, I want to participate and try to develop a neural network in charge of categorizing each clothing it sees. Therefore, we need to use deep learning and understand how it works. The concept of deep learning started in 1943 when Warren McCulloch and Walter Pitts created a computer model based on the human brain’s neural networks. They used a combination of mathematics and threshold logic algorithms to mimic the thought process. Since then, deep learning has evolved steadily, with two significant developmental breaks over the years. The progress of the basics of a continuous Back Propagation Model by Henry J. Kelley in 1960, and when Stuart Dreyfus came up with a simpler version based only on the chain rule in 1962. Now, it is an important topic. Scientists use deep learning algorithms with multiple processing layers to make better models capable of understanding large quantities of unlabelled data, such as photos with no description, voice recordings, etc. We want to use those algorithms, especially Object Detection and Semantic Segmentation, to start a project where we want to detect different pieces of clothing in a large dataset established by Fashion RGB Images. The purpose is to carry out a study on both implementations. We want to train a model several times with different parameters and datasets, trying to achieve the most optimal results from both. Once we have the results, we will compare them to see which is best for our case and study why it is the best.