Visualizing and Validating Learned Image Embeddings
Open Access DepositedImage embeddings have become a core representational tool in Computer Vision, vital for image retrieval and recognition in domains with millions of categories, and integral to the translations between image and language. Currently, learning these embeddings is based on Deep Learning architecture trained to map images into a space where semantically similar images are mapped to similar locations and semantically different images are mapped to distant locations. It is not often clear exactly what features are extracted from the images when doing this mapping, and work on visualization tools in this domain is very limited. This proposal asks four questions to improve these visualizations. First, how could we used such visualization? Second, how can we better visualize the structure of the mapping? Third, can we better show which parts of images are related to each other when images are judged to be similar or dissimilar? Fourth, for multi-modal embeddings that relate text and imagery, can we measure the representational consistency between images and text embeddings to predict text-driven image classification?
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.