We found that the best performing model was generated when we initially trained the full network on the classification task. It took on average about one day to train a model. Two hundred images from each class were held out from the training data and used for testing. A better organization of the model would likely have been to split the tasks into two separate classification tasks, one binary (handbag, no handbag) and if a handbag is present, a multiclass brand classifier, however in this prototype we made it a one level classifier. We used the Inception-v3 architecture and this model which we initialized from a model pre-trained on the ImageNet dataset available here. This machine is equipped with a single Tesla K80 gpu. The model was implemented in TensorFlow running on an AWS p2.xlarge instance. In total we collected a relatively balanced dataset of approximately 17,000 images across these 7 brands and 1 negative class:Īs we iterated on training the model we used the current version of the model to ‘pre-tag’ the images as this greatly sped up the data collection process and validation. An image was allowed to contain more than one handbag but since we did not include any object detection we only included multiple handbags if they were the same brand. The data contains selfies and other amateur images, white background studio style images, professional fashion and runway images. All the images were reviewed manually before being added to the dataset. The data used here was collected from Instagram using both hashtags as well as brand and fan pages. If you’re not yet familiar with neural networks or general Machine Learning terminology, take a look at our Neural Network Primer first. Indeed, a “human expert” can make a reasonably good prediction of the handbag's brand without having seen that exact model. These features range from the more obvious (e.g., patterns, logos) to the less visible (e.g., textures, pockets, latches, straps). Many brands have features that distinguish them visually. Furthermore, from a computer vision perspective, handbags are rather complex objects. We decided to focus on handbags because they are objects from the fashion domain, where Condé Nast already has a significant presence. These graphical media provide exciting frontiers where we can utilize and Machine Learning to enhance our experiences.Īfter considering a few ideas, we decided to prototype a handbag brand classifier. In addition to textual content, our stories include vibrant videos and images. This information serves as useful building blocks for other tools to improve both the user and editorial experience. (e.g., people, organizations, products), and keywords in an article published by any of our brands. These efforts have mainly focused on Natural Language Processing (NLP) where we have created tools that can automatically detect topics, entities Over the past few years, we here at Condé Nast have invested heavily in building Machine Learning (ML) tools to help us understand our content and how our users interact with it. For a primer on Neural Network concepts, please visit our first post in this series.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |