| Popis: |
eTryon’s WP3 is centered around building systems for 1) pattern recognition in fashionimagery, 2) fashion trend forecasting and garment popularity prediction and 3) fashionrecommendations. The current deliverable (D3.2) uses the fashion labels and visual features extracted by models built in the framework of D3.1, in order to develop apipeline capable of identifying fashion trends and predicting the popularity of newgarment designs for different market segments, based on age and gender. This deliverable describes in detail the research work carried out by CERTH andMallzee, including the data collection process, the training and evaluation of machinelearning models for carrying through the tasks of fashion trend forecasting and themarket-segmented garment popularity prediction on new garment designs. We proposed a hybrid quasi auto-regressive architecture, H-QAR, which combines amulti-layer perceptron utilizing the visual and categorical features of a garment and anauto-regressive network module utilizing the time series of the garment’s categories and attributes. H-QAR enables inference on new garment designs based on the garment’svisual appearance as well as the garment’s category and attribute time series in order tocompensate for the lack of historical data. Our proposed architecture was able to surpass the baseline models in both the Mallzee dataset and on a fashion benchmarkdataset SHIFT15m. Moreover, our experiments on SHIFT15m surpassed the SotA onthe task of outfit popularity prediction, and similarly, our experiments on the Paris toBerlin dataset were able to surpass the SotA for the task of fashion trend forecasting. The developed models will be deployed through an API, here elaborated in Section 6,that will be used in the Designer app and as input features for the recommender systemsin the Influencer and the e-commerce apps. The designers will be able to upload an image of a new garment design, the image will be passed through the “computer visionAPI” (from D3.1). The visual features along with the predicted category and attributes willbe passed to the “popularity prediction API” which will predict the popularity score of thegiven garment design. |
