Virtual fitting room

A virtual fitting room (VFR) for a network of European multi-brand retailers selects clothes by superimposing items of clothing onto a customer photo. Errors in fitting should be kept to a minimum, as inaccuracies may mean the return of goods and lost sales.

Various online fitting programs exist already. Their major downside is that you have to enter too much information about yourself. Besides, the quality of the fitting leaves much to be desired. The distinctive feature of the VFR is that most of the model's parameters are identified automatically from the photograph.

VFR offers:

• Identification of model's parameters from the customer photograph;
• Selection of the model and size of the chosen item based on the attributes of the model;
• Selection and visualisation of full outfits (of various brands if desired) taking into account the customer's preferences (casual, business style, evening wear, etc.);
• Fitting of clothes with the help of a mobile device, online store or stationary fitting room;
• Self-service in store;
• Ordering clothes without visiting a store;
• Faster fitting and less manual work;
• Less workload on consultants and merchandisers and partial reduction of personnel as a result;
• Sales growth;
• Captivation of clientele through technological innovation;
• Increased competitiveness through the use of technology.

Structure

Body model

A customer looks at the screen, which displays a request to take a full-length portrait photograph. The screen displays instructions and markers indicating the desired perimeters of the body in order to obtain an image that can be later calculated by the algorithm.

When a person enters the video zone, the algorithm starts analysing the in-built webcam stream. The photograph is taken when the customer's body is in the correct pose and position relative to the lens.

The photo is then sent to a server where the model's contours are identified and a two-dimensional vector model is built.

1. A photo of a human body is fed to the input. The photo should show the person against a contrasting background, and the clothes the person is wearing should be tight-fitting and preferably solid-coloured. Other people or objects are not allowed in the frame.
2. The body contour is calculated using the OpenCV library.
3. The resulting contour is validated. The person in the photo should be standing straight, the body should be positioned symmetrically, the legs should be shoulder-width apart, the feet should be facing forward, and the arms should be spread at a 45° angle from the body with the palms inward.
4. The bends of the contour are used to determine the coordinates of the nodal points.
5. The distances in pixels between the nodal points of the contour are calculated horizontally and vertically.
6. Typical models for a man and a woman (Slava and Coco) are used as basic vector models.
7. The correction of the nodal points along the coordinate axes (delta) is calculated to adjust the basic vector model for a man or woman.
8. As the output we have an array of coordinates for the nodal points of the vector model.

Model adjustme

The two-dimensional model is displayed to the user in the form of a monochrome, semi-transparent filled contour superimposed over the original photo. There are tools for scaling the photos relative to the model for a more accurate fit. The contour of the model has nodal points, which can be moved by the user within set limits. Correction boundaries are set for every point.

The following nodal points are given.

Moving a nodal point leads to a mirrored movement of the symmetrical point relative to the vertical axis. When capturing a nodal point, graphic elements are displayed that indicate acceptable directions and correction limits.

A user can specify one or more parameters: height, waist circumference etc. The other parameters are calculated automatically taking into account the proportions of various parts of the body and the anatomical characteristics of the human body in general. Since a two-dimensional projection is used, the calculations of circumferential measurements are inaccurate and may require manual adjustment of the parameters.

The model's parameters and an image set containing the model and the photograph are saved in the user's profile.

Fitting clothes

The catalogue is a standard online store interface with search and product view functionality, allowing customers to see the cost and availability of items. A key feature is the hint function, which shows the availability of clothes that are suitable in size and match the customer's desired style.

The selected clothes and accessories are combined into an ‘image’ that is superimposed on the body model. The clothes can be found at a stationary fitting room by scanning the relevant barcode.

The fitting screen displays the body model and, if desired, the customer's photo. A catalogue of goods of a specific type is opened by clicking on an ‘unclothed’ part of the model. A new item of clothing is placed on the corresponding body part. The size is calculated automatically, taking into account the compatibility of the new items with the previously selected ones.

The customer can use the “Share photo” button with the superimposed items from the image. The resulting image can be transferred from the mobile application to a stationary fitting room and vice versa using a QR code.

2D clothing models

Shop assistants or brand merchandisers can prepare virtual models of every item of clothing in advance.

1. The item is photographed against a contrasting background.
2. The item should be placed on a mannequin whose colour matches the background.
3. The size of the mannequin should correspond to the size of the clothing item.
4. The background is removed from the photo.
5. 2D models are created for every size and colour of the item of clothing.
6. It is impossible to extrapolate the proportions onto various sizes of one model. For example, in two dresses of different sizes, the ratio of the sleeve lengths to the waist circumference may not coincide.
7. Manual measurements are taken for every model of every item of clothing.

Outlook

A working version with 2D static body and clothing models was implemented. In future we plan to implement 3D models and augmented reality which will enable video processing from the camera facing the customer, allowing for instant fitting.