What is it about?
First, it is very hard and time-consuming to manually create outfits datasets. Therefore, we aim to reduce the dependency on a large labeled data and effectively use unlabeled fashion images for learning outfit compatibility. Second, since existing datasets are small in size, we propose a semi-automatic way for curating large-scale dataset by applying association mining principles to customer transaction history.
Featured Image
Photo by Heather Ford on Unsplash
Why is it important?
We find that our proposed method is effective for learning fashion compatibility even when we access only 5% of the data. Results demonstrate that our semi-supervised approach is on par with many fully-supervised approaches. Further, we include details on the data collection strategy for Fashion Outfits dataset.
Perspectives
In this work, we used a small fraction of annotated dataset and were able to achieve significant performance on standard benchmarks. I am hoping that this work will provide inspiration for new direction of works that including few-shot learning and self-supervised learning for fashion compatibility.
Ambareesh Revanur
Carnegie Mellon University
Read the Original
This page is a summary of: Semi-Supervised Visual Representation Learning for Fashion Compatibility, September 2021, ACM (Association for Computing Machinery),
DOI: 10.1145/3460231.3474233.
You can read the full text:
Resources
Contributors
The following have contributed to this page
