GeoStyle: Discovering Fashion Trends and Events

Posted 2020-11-23 cyd1310997

-	-
GeoStyle: Discovering Fashion Trends and Events---ICCV 2019 Open Access Repository

GeoStyle: Discovering Fashion Trends and Events
GeoStyle:发现时尚趋势和事件

Abstract
摘要

Understanding fashion styles and trends is of great potential interest to retailers and consumers alike.
了解时尚风格和趋势对零售商和消费者都有很大的潜在利益。
The photospeople upload to social media are a historical and publicdata source of how people dress across the world and atdifferent times.
人们上传到社交媒体上的照片是一个历史和公共数据来源，记录了世界各地不同时期人们的着装。
While we now have tools to automaticallyrecognize the clothing and style attributes of what peopleare wearing in these photographs, we lack the ability to analyze spatial and temporal trends in these attributes or makepredictions about the future.
虽然我们现在有工具自动识别人们在这些照片中所穿的衣服和风格，但我们缺乏分析这些属性的时空趋势或预测未来的能力。
In this paper we address thisneed by providing an automatic framework that analyzeslarge corpora of street imagery to (a) discover and forecast longterm trends of various fashion attributes as wellas automatically discovered styles, and (b) identify spatiotemporally localized events that affect what people wear.
在本文中，我们通过提供一个自动的框架来解决这一需求，该框架通过分析街道图像的大型语料库(a)发现和预测各种时尚属性的长期趋势以及自动发现的风格，(b)识别影响人们穿着的时空本地化事件。
Weshow that our framework makes long term trend forecaststhat are > 20% more accurate than prior art, and identifieshundreds of socially meaningful events that impact fashionacross the globe.
我们认为，我们的框架使长期趋势预测比现有技术准确20%，并确定了数百个影响全球时尚的有社会意义的事件。

1. Introduction
2. 介绍

Each day, we collectively upload to social media platforms billions of photographs that capture a wide range ofhuman life and activities around the world.
每天，我们都要向社交媒体平台上传数十亿张照片，这些照片记录了世界各地的人类生活和活动。
At the same time,object detection, semantic segmentation, and visual searchare seeing rapid advances 13 and are being deployed atscale 22.
与此同时，目标检测、语义分割和视觉搜索正在迅速发展，并在第22级部署。
With largescale recognition available as a fundamental tool in our vision toolbox, it is now possible to askquestions about how people dress, eat, and group across theworld and over time.
随着大规模的认可成为我们愿景工具箱中的基本工具，现在我们可以询问关于世界各地人们如何穿衣、饮食和群体的问题。
In this paper we focus on how people dress.
在这篇论文中，我们关注的是人们的穿着。
In particular, we ask can we detect and predictfashion trends and styles over space and timeWe answer these questions by designing an automatedmethod to characterize and predict seasonal and yearoveryear fashion trends, detect social events (e.g., festivals orsporting events) that impact how people dress, and identify socialeventspecific style elements that epitomize theseevents.
特别是,我们问我们能检测和predictfashion趋势和风格在空间和设计一个automatedmethod timeWe回答这些问题的描述和预测季节性和yearoveryear时尚趋势,发现社会事件(例如,节日orsporting事件)影响人们如何着装,并确定socialeventspecific风格元素概括theseevents。
Our approach uses existing recognition algorithmsto identify a coarse set of fashion attributes in a large corpusof images.
我们的方法使用现有的识别算法来识别大量图像中的一组粗糙的时尚属性。
We then fit interpretable parametric models oflongterm temporal trends to these fashion attributes.
然后，我们将长期的时间趋势的可解释参数模型拟合到这些时尚属性中。
Thesemodels capture both seasonal cycles as well as changes inpopularity over time.
这些表情捕捉了季节周期和流行度随时间的变化。
These models not only help in understanding existing trends, but can also make up to 20% moreaccurate, temporally finegrained forecasts across long timescales compared to prior methods 1.
这些模型不仅有助于理解现有的趋势，而且与以前的方法相比，它们还可以在较长时间尺度上提高20%的准确性和时间上的细粒度预测。
For example, we findthat yearonyear more people are wearing black, but thatthey tend to do so more in the winter than in the summer.
例如，我们发现一年比一年多的人穿黑色衣服，但是他们在冬天穿的比夏天多。
Our framework not only models longterm trends, but alsoidentifies sudden, shortterm changes in popularity that buckthese trends.
我们的框架不仅模拟了长期的趋势，还识别出了突然的、短期的、与这些趋势相悖的流行变化。
We find that these outliers often correspondto festivals, sporting events, or other large social gatherings.
我们发现，这些异常值通常与节日、体育赛事或其他大型社交聚会相对应。
We provide a methodology to automatically discover theevents underlying such outliers by looking at associated image tags and captions, thus tying visual analysis to textbaseddiscovery.
我们提供了一种方法，通过查看相关的图像标记和标题，自动发现这些异常值背后的事件，从而将可视化分析与基于文本的发现结合起来。
We find that our framework finds understandablereasons for all of the most salient events it discovers, and inso doing surfaces intriguing social events around the worldthat were unknown to the authors.
我们发现，我们的框架为它所发现的所有最显著的事件找到了可理解的原因，并在此过程中揭示了作者不知道的世界各地有趣的社会事件。
For example, it discoversan unusual increase in the color yellow in Bangkok in earlyDecember, and associates it with the words father, day,king, live, and dad.
例如，它发现12月初曼谷的黄色有所增加，并将其与“父亲”、“日子”、“国王”、“生活”和“爸爸”等词联系起来。
This corresponds to the kingsbirthday, celebrated as Fathers Day in Thailand by wearing yellow 36.
这一天正好是国王的生日，在泰国，人们穿着黄色衣服庆祝父亲节。
Our framework similarly surfaces eventsin Ukraine (Vyshyvanka Day), Indonesia (Batik Day), andJapan (Golden Week).
我们的框架同样涉及乌克兰(Vyshyvanka日)、印度尼西亚(Batik日)和日本(黄金周)的事件。
Figure 1 shows more of the worldwide events discovered by our framework and the clothesthat people wear during those events.
图1显示了我们的框架发现的更多的全球事件，以及人们在这些事件中穿的衣服。
We further show that we can predict trends and eventsnot just at the level of individual fashion attributes (such aswearing yellow), but also at the level of styles consistingof recurring visual ensembles.
我们进一步证明，我们不仅可以在个人时尚属性(如穿黄色衣服)的层面上预测趋势和事件，还可以在反复出现的视觉效果的风格层面上预测趋势和事件。
These styles are identifiedby clustering photographs in feature space to reveal styleclusters clusters of people dressed in a similar style.
这些风格是通过在特征空间中聚集照片来识别的，以揭示风格爱好者聚集在一起穿着相似风格的人。
Ourforecasts of the future popularity of styles are just as accurateas our predictions of individual attributes.
我们对未来流行款式的预测和对个人属性的预测一样准确。
Further, we canrun the same event detection framework described above onstyle trends, allowing us to not only automatically detectsocial events, but also associate each event with its owndistinctive style a stylistic signature for each event.
此外，我们还可以在style trends上运行与上述相同的事件检测框架，不仅可以自动检测社会事件，还可以将每个事件与它自己独特的风格关联起来，即每个事件的风格特征。
Our contributions, highlighted in Figure 2, include We present an automated framework for analyzing thetemporal behavior of fashion elements across the globe.
我们的贡献(在图2中突出显示)包括我们提供了一个自动化框架，用于分析全球各地的时尚元素的临时行为。
Our framework models and forecasts longterm trendsand seasonal behaviors.
我们的框架模型和预测长期趋势和季节行为。
It also automatically identifiesshortterm spikes caused by events like festivals andsporting events.
它还会自动识别由节日和体育赛事等事件造成的短期峰值。
Our framework automatically discovers the reasons behind these events by leveraging textual descriptions andcaptions.
我们的框架通过利用文本描述和说明自动发现这些事件背后的原因。
We connect events with signature styles by performingthis analysis on automatically discovered style clusters.
通过对自动发现的样式集群进行分析，我们将事件与签名样式联系起来。