A Comparative Analysis of Practices in Training Deep Models for Fashion Attribute Detection

With the rapid increase of smartphone technologies and social media apps, we live in a time where every day billions of photographs are shared by people through their personal devices, and a large amount of these photos involves person images or selfies. In this study, we investigate the problem of recognizing and classifying fashion attributes in person images. We perform extensive experiments on the StreetStyle-27k dataset with the, a recently proposed large-scale dataset collected for this purpose, in which we analyze the current best practices such as Sthochasthic Gradient Descent with Warm Restarts,Focal loss,Temperature Scaling that are generally used for effective training of deep convolutional networks. Especially, we elaborate on a specific challenge that commonly arise in in-the-wild problems such as ours, which is learning when the distribution of labels is unbalanced.The results we get with the best model is %3.67 better than StreetStyle. We hope that our results will shed some light and be useful to other researchers.