Saved in:
Bibliographic Details
Main Authors: Zhu, Xinliang, Huang, Michael, Ding, Han, Yang, Jinyu, Chen, Kelvin, Zhou, Tao, Neiman, Tal, Xie, Ouye, Tran, Son, Yao, Benjamin, Gray, Doug, Bindal, Anuj, Dhua, Arnab
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2412.13364
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912160246923264
author Zhu, Xinliang
Huang, Michael
Ding, Han
Yang, Jinyu
Chen, Kelvin
Zhou, Tao
Neiman, Tal
Xie, Ouye
Tran, Son
Yao, Benjamin
Gray, Doug
Bindal, Anuj
Dhua, Arnab
author_facet Zhu, Xinliang
Huang, Michael
Ding, Han
Yang, Jinyu
Chen, Kelvin
Zhou, Tao
Neiman, Tal
Xie, Ouye
Tran, Son
Yao, Benjamin
Gray, Doug
Bindal, Anuj
Dhua, Arnab
contents Image to image matching has been well studied in the computer vision community. Previous studies mainly focus on training a deep metric learning model matching visual patterns between the query image and gallery images. In this study, we show that pure image-to-image matching suffers from false positives caused by matching to local visual patterns. To alleviate this issue, we propose to leverage recent advances in vision-language pretraining research. Specifically, we introduce additional image-text alignment losses into deep metric learning, which serve as constraints to the image-to-image matching loss. With additional alignments between the text (e.g., product title) and image pairs, the model can learn concepts from both modalities explicitly, which avoids matching low-level visual features. We progressively develop two variants, a 3-tower and a 4-tower model, where the latter takes one more short text query input. Through extensive experiments, we show that this change leads to a substantial improvement to the image to image matching problem. We further leveraged this model for multimodal search, which takes both image and reformulation text queries to improve search quality. Both offline and online experiments show strong improvements on the main metrics. Specifically, we see 4.95% relative improvement on image matching click through rate with the 3-tower model and 1.13% further improvement from the 4-tower model.
format Preprint
id arxiv_https___arxiv_org_abs_2412_13364
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Bringing Multimodality to Amazon Visual Search System
Zhu, Xinliang
Huang, Michael
Ding, Han
Yang, Jinyu
Chen, Kelvin
Zhou, Tao
Neiman, Tal
Xie, Ouye
Tran, Son
Yao, Benjamin
Gray, Doug
Bindal, Anuj
Dhua, Arnab
Computer Vision and Pattern Recognition
Image to image matching has been well studied in the computer vision community. Previous studies mainly focus on training a deep metric learning model matching visual patterns between the query image and gallery images. In this study, we show that pure image-to-image matching suffers from false positives caused by matching to local visual patterns. To alleviate this issue, we propose to leverage recent advances in vision-language pretraining research. Specifically, we introduce additional image-text alignment losses into deep metric learning, which serve as constraints to the image-to-image matching loss. With additional alignments between the text (e.g., product title) and image pairs, the model can learn concepts from both modalities explicitly, which avoids matching low-level visual features. We progressively develop two variants, a 3-tower and a 4-tower model, where the latter takes one more short text query input. Through extensive experiments, we show that this change leads to a substantial improvement to the image to image matching problem. We further leveraged this model for multimodal search, which takes both image and reformulation text queries to improve search quality. Both offline and online experiments show strong improvements on the main metrics. Specifically, we see 4.95% relative improvement on image matching click through rate with the 3-tower model and 1.13% further improvement from the 4-tower model.
title Bringing Multimodality to Amazon Visual Search System
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2412.13364