Embedding the CDN Framework in Online Fashion Style Databases
A Dataset in fashion research is the collation of data points about fashion styles or fashions and it includes the following elements: examples of garments that might be worn, what these garments might be called upon by men and women alike, a list of the most popular colors and other things of interest. Many of the items on the datasets will be from a particular decade or season. The reason for this is to provide a basis upon which to draw comparisons between styles and seasons. As an example, if we looked at datasets about Fifth Generation style in which jeans were predominant in the 1980s, then it is easy to see that this was a longer looking era than many of the earlier decades. Also, the styles associated with this decade were very distinct. A similar case can be made when we look at datasets about current fashion styles, and in which the most common items of clothing are casual trousers, work boots, skirts, and tops.
The creation of a good quality Dataset can be achieved in two ways: either using traditional classification methods such as the Categorization Review (CR), the Review of Fashion Data (RFD), the RCP, or the SMART method, or using pre-trained CD-NMS data. Each of these approaches requires the compilation and storage of large amounts of unlabeled data. In addition, once all the data has been collected it is important to perform a significant number of classifications. For instance, if the Dataset consists of only one hundred examples of a particular clothing item, it is relatively easy to identify the twenty best examples and make a quick decision about which groupings would best fit the data.
The use of pre-trained CD-NMS data is based on the idea that a single data set can be used to predict and forecast future fashion trends, especially when the ensemble of items is consistent with previously known fashion trends. For example, if the Dataset consists of one hundred examples of women’s jeans, it is quite feasible to infer that in the next year a majority of these jeans will have slender legs and/or larger cup sizes. However, given that there are no other comparable data sets available, the use of this Dataset in fashion trend forecasting is somewhat limited.
Both approaches, the traditional classification methods and the pre-trained CD-NMS sets, rely on the knowledge of human body characteristics such as height, weight, eye and hair color, skin type, and other physical features. This knowledge is often difficult to collect in a real-time fashion style context since different human bodies often conform to widely varying physical traits at different times. In addition, it is difficult to extract reliable parameters for each human attribute since it is not uncommon for human body characteristics to change for a myriad of reasons.
As an alternative to traditional classifications, the Dataset can also employ Machine Learning to predict future styles. This includes supervised learning and reinforcement techniques. Once enough consistent patterns of behavior is accumulated, the machine can automatically generalize the known trends from its training data. The CD-Nsns allows for the creation of trend data with a high degree of confidence, since the supervised learning approach takes into account a wide range of inputs and outputs ensuring a high level of statistical significance. While the latter reduces the range of possible classifications from human classifications, it still relies heavily on the knowledge of the physical attributes of the human body which are inherently stochastic in nature.
Once sufficient input images and/or videos have been generated, the CD-NMS can then attempt to identify relevant components for each attribute. Based on the pre-training data, the machine can then try to associate each input image or video with a label associated to that image or video. This allows the system to create classifications from a large, high volume input. Of course, this process relies heavily on the knowledge of the human body, but it has the advantage of being able to generate high quality classifications even from small, poorly specified feature images. Thus, it is an important component in the toolbox of any good fashion style classification software.