Jointly End-To-End Embedding Learning for Large-Scale Sticker Similarity Search

Vinhhiep Le
LINE Fukuoka Data Labs Machine Learning Engineer
https://linedevday.linecorp.com/jp/2019/sessions/S1-03

Transcript

1. 2019 DevDay Jointly End-To-End Embedding Learning for Large-Scale Sticker Similarity Search > Vinhhiep Le > LINE Fukuoka Data Labs Machine Learning Engineer
2. Agenda > Sticker similarity definition > Common approach and problem > Solution and experiment > Summary
3. Sticker Similarity Search > Content-based image retrieval: Given a sticker, retrieve a list of similar stickers.
4. Common Approach and Problem > Represent images by embedding and retrieve similar images by ranking similarity scores using a distance metric > Problem: When it comes to hundred of millions stickers, exhaustive searching the whole database is computationally infeasible
5. Solution > Multiple levels of filters:
6. Cheap Representations > Filter out hopeless candidates through multiple filter layers • Learning: binary code • Non-learning: Bag of Visual Words (BOVW)
7. Cheap Representations > Why binary code • Could be represented as a document • Fast distance calculation using Hamming distance
8. Our Approach > Three-level Filters: > We jointly learn binary code and float embedding in on network end-to-end!
9. Experimental Results > Test set: 14,440 stickers, 480 packages, each package has 30 similar stickers. > Evaluation metric: Precision@TopK
10. Summary > Multi-layer search is good to deal with large-scale search > Represent the image by learning method is better than non-learning
11. Thank You