Machine Learning - FashionTagger project
We automatically recognize and tag clothes
play video
FashionTagger cleverly recognizes a kind of clothing and assigns right attributes e.g. colour, neckline, pattern. E-commerce customers can now quickly and efficiently search clothes and shop online.
How does it works?
Step 1
FashionTagger software automatically detects a model in the picture. The algorithm finds the position of various clothes on the picture for further analysis.
Step 2
Recognized clothing items are send through classification process, which assigns attributes such as color, pattern neckline. Each type of clothing has its own characteristic attributes, which serve customers to effectively search and shop online.
Step 3
Thanks to methods implemented in FashionTagger, the customer gets precise location of a piece of clothing at the picture. Our system provides robust information regarding any features of detected clothes.
FashionTagger software uses algorithms based on Deep Learning methods. The way the system works could be described in following steps (1) pointing the input graphic file - photo of a model, (2) recognizing of the human being at the photo, (3) recognizing each piece of clothing - trousers, shoes, purse etc. (4) extracting features that are characteristic for each clothing, (5) returning a list of items containing data regarding searched products for which similarity coefficient has the highest value.
CTA's solution extracts from the picture every part consistent of similar color and form. At the same time the software is robust to elements of low importance such as background. For each separated part it is necessary to find its distinctive attributes and to use transformation allowing quick search for similar items in the database. The process of comparison includes information about color, pattern, shape and style of the item. To achieve this goal we are using Deep Learning methods alongside Convolutional Neural Network which are for now the greatest tools for automatic image analysis. High capability of the system was achieved due to using graphic cards processors.
object recognition variant
increasingly complex features
simple inputs
Live demo