В Wildberries провели воспитательную беседу с менеджером пункта выдачи заказов, который угрожал клиентам в ХМАО. Об этом URA.RU сообщили в пресс-службе компании.

Wildberries сделала заявление из-за менеджера, который угрожал клиентам в ХМАО

В Нефтеюганске (ХМАО) на дверях пункта выдачи заказов Wildberries появились записки с угрозами ограничить доступ к заказу вещей на точку, если кто-то решит поставить ...

В Нефтеюганске на дверях пункта выдачи Wildberries появились угрозы клиентам: фото

9 декабря, Минск /Корр. БЕЛТА/. Началась подготовка земельного участка в "Великом камне" под распределительный центр Wildberries. Об этом рассказал первый заместитель ...

Началась подготовка земельного участка в "Великом камне" под распределительный центр Wildberries

На тюменских пунктах выдачи заказов Wildberries начался предновогодний ажиотаж. Об этом сообщили горожане, обратившиеся в редакции. По их словам, на сервисе увелич...

Wildberries в Тюмени: на ПВЗ начался предновогодний ажиотаж

Стоимость возврата товаров на Wildberries для жителей ХМАО может варьироваться от расстояния обратной доставки, сообщили URA.RU представители маркетплейса.

Возврат товаров на Wildberries: маркетплейс объяснил причину высоких цен для жителей ХМАО

Маркетплейс Wildberries начал скрывать негативные отзывы, которые не относятся к товару или продавцу. Такие отзывы не будут учитываться при расчёте рейтинга товара ...

Wildberries стал скрывать негативные отзывы, если они не про товар или продавца

Первым клиентом Wildberries стала жительница Нефтеюганска (ХМАО). Об этом заявила гендиректор Wildberries Татьяна Бакальчук во время встречи с губернатором региона Натальей ...

Первым клиентом Wildberries стала жительница Нефтеюганска, - Татьяна Бакальчук

Wildberries will build a logistics hub in Chita for a planned 7 billion rubles ($71), the press service of the Trans-Baikal Territory Development Co

Wildberries Will Build 7-Bln Ruble Logistical Hub In Chita - Authorities

CHITA. Oct 18 (Interfax) - Wildberries will build a logistics hub in Chita for a planned 7 billion rubles, the press service of the Trans-Baikal Territory Development Corporation said. "The Trans-Baikal Territory Development Corporation has been working ...

Wildberries will build 7-bln ruble logistical hub in Chita - authorities

Humboldt Roller Derby is gearing up for its next doubleheader set for Saturday at Redwood Acres. The event is Halloween-themed.

Roller derby | Humboldt squad’s next doubleheader set for Saturday

Wildberries

wildberries

The Evolution of Data Center Cooling: From Air-Based Methods to Free Cooling

Die evolusie van datasentrumverkoeling: van luggebaseerde metodes tot gratis verkoeling

የውሂብ ማእከል ማቀዝቀዝ ዝግመተ ለውጥ፡- ከአየር-ተኮር ዘዴዎች ወደ ነፃ ማቀዝቀዝ

تطور تبريد مراكز البيانات: من الأساليب المعتمدة على الهواء إلى التبريد المجاني

Uka Evolución del Centro de Datos de Cooling: Métodos Basados en Aire ukanakat Libre Cooling ukar puriñkama

Məlumat Mərkəzinin Soyutmasının Təkamülü: Hava Əsaslı Metodlardan Sərbəst Soyutmaya

Эвалюцыя астуджэння цэнтра апрацоўкі дадзеных: ад паветраных метадаў да вольнага астуджэння

Еволюцията на охлаждането на центъра за данни: от въздушни методи до свободно охлаждане

Evolucija hlađenja podatkovnih centara: od metoda na bazi zraka do slobodnog hlađenja

L'evolució de la refrigeració del centre de dades: dels mètodes basats en l'aire a la refrigeració lliure

Evoluce chlazení datových center: od vzduchových metod k volnému chlazení

Udviklingen af datacenterkøling: Fra luftbaserede metoder til frikøling

Η Εξέλιξη της Ψύξης του Κέντρου Δεδομένων: Από τις Μεθόδους Με βάση τον Αέρα στην Ελεύθερη Ψύξη

Datu-zentroen hoztearen bilakaera: airean oinarritutako metodoetatik free hozteraino

سیر تکامل خنک‌سازی مرکز داده: از روش‌های مبتنی بر هوا تا خنک‌سازی رایگان

Palvelinkeskusten jäähdytyksen kehitys: ilmapohjaisista menetelmistä ilmaiseen jäähdytykseen

A evolución do arrefriamento do centro de datos: dos métodos baseados en aire ao free cooling

האבולוציה של קירור מרכזי נתונים: משיטות מבוססות אוויר לקירור חופשי

Evolucija hlađenja podatkovnog centra: od zračnih metoda do slobodnog hlađenja

Evolisyon nan refwadisman sant done: Soti nan metòd ki baze sou lè rive nan refwadisman gratis

Az adatközponti hűtés fejlődése: a levegő alapú módszerektől a szabad hűtésig

Տվյալների կենտրոնի սառեցման էվոլյուցիան. օդի վրա հիմնված մեթոդներից մինչև անվճար սառեցում

Evolusi Pendinginan Pusat Data: Dari Metode Berbasis Udara ke Pendinginan Gratis

L'evoluzione del raffreddamento dei data center: dai metodi basati sull'aria al raffreddamento gratuito

მონაცემთა ცენტრის გაგრილების ევოლუცია: ჰაერზე დაფუძნებული მეთოდებიდან თავისუფალ გაგრილებამდე

Деректер орталығын салқындату эволюциясы: ауа негізіндегі әдістерден еркін салқындату

ការវិវឌ្ឍន៍នៃការធ្វើឱ្យត្រជាក់នៃមជ្ឈមណ្ឌលទិន្នន័យ៖ ពីវិធីសាស្ត្រផ្អែកលើខ្យល់ទៅភាពត្រជាក់ដោយឥតគិតថ្លៃ

Маалымат борборунун муздатуусунун эволюциясы: абага негизделген ыкмалардан эркин муздатууга

Evolution ya Refroidissement ya Centre de Données: Kobanda na ba méthodes basées na mopepe tii na refroidissement gratuit

ການວິວັດທະນາການຂອງສູນຂໍ້ມູນຄວາມເຢັນ: ຈາກວິທີການທີ່ອີງໃສ່ອາກາດໄປສູ່ຄວາມເຢັນຟຣີ

Duomenų centrų aušinimo evoliucija: nuo oru naudojamų metodų iki nemokamo aušinimo

Datu centra dzesēšanas evolūcija: no gaisa metodēm līdz bezmaksas dzesēšanai

Ny fivoaran'ny fampangatsiahana ivotoerana data: avy amin'ny fomba mifototra amin'ny rivotra mankany amin'ny fangatsiahana maimaim-poana

Еволуцијата на ладењето на центарот за податоци: од методи базирани на воздух до бесплатно ладење

Дата төвийн хөргөлтийн хувьсал: Агаарт суурилсан аргуудаас чөлөөт хөргөлт хүртэл

Evolusi Penyejukan Pusat Data: Daripada Kaedah Berasaskan Udara kepada Penyejukan Percuma

डाटा सेन्टर कूलिङको विकास: वायुमा आधारित विधिहरूबाट नि:शुल्क कूलिङसम्म

De evolutie van datacenterkoeling: van luchtgebaseerde methoden naar vrije koeling

Tlhagelelo ya go Fodiša ga Setsi sa Datha: Go Tšwa go Mekgwa ye e Theilwego Moyeng go ya go go Fodiša ga Mahala

Ewolucja chłodzenia centrów danych: od metod opartych na powietrzu do chłodzenia swobodnego

د ډیټا سنټر یخولو تکامل: د هوا پر اساس میتودونو څخه وړیا یخولو ته

Willaypa chawpinpi q'uñichiypa wiñaynin: Wayrapi ruwasqa ruwaykunamanta mana qullqiyuq q'uñichiyman

Evoluția răcirii centrului de date: de la metode pe bază de aer la răcire liberă

Ubwihindurize bwa Data Centre Cooling: Kuva Muburyo bushingiye ku kirere kugeza ubukonje bwubusa

දත්ත මධ්‍යස්ථාන සිසිලනයේ පරිණාමය: වායු පාදක ක්‍රමවල සිට නොමිලේ සිසිලනය දක්වා

Evolúcia chladenia dátových centier: od metód založených na vzduchu po voľné chladenie

Iko Evolution yeData Center Kutonhodza: Kubva kuMhepo-Yakavakirwa Nzira kuenda Kumahara Kutonhodza

Kobcinta Qaboojinta Xarunta Xogta: Laga soo bilaabo Hababka Hawada Ku Salaysan ilaa Qaboojinta Bilaashka ah

Evolucioni i ftohjes së qendrës së të dhënave: Nga metodat e bazuara në ajër te ftohja falas

Еволуција хлађења центара података: од ваздушних метода до бесплатног хлађења

Utvecklingen av kylning av datacenter: från luftbaserade metoder till fri kylning

Mageuzi ya Kupoeza kwa Kituo cha Data: Kutoka Mbinu Zinazotegemea Hewa hadi Kupoeza Bila Malipo

தரவு மைய குளிரூட்டலின் பரிணாமம்: காற்று அடிப்படையிலான முறைகளிலிருந்து இலவச குளிரூட்டலுக்கு

Эволютсияи сардшавии маркази додаҳо: аз усулҳои ҳавоӣ то сардшавии ройгон

วิวัฒนาการของระบบทำความเย็นศูนย์ข้อมูล: จากวิธีการใช้ลมสู่ระบบทำความเย็นแบบอิสระ

ፍልቀት መዝሓሊ ማእከል ዳታ፡ ካብ ኣየር መሰረት ዝገበረ ሜላታት ናብ ነጻ ምዝሓል

Maglumat merkeziniň sowadyşynyň ewolýusiýasy: Howa esasly usullardan mugt sowatmaga çenli

Ang Ebolusyon ng Data Center Cooling: Mula sa Air-Based Methods hanggang sa Libreng Paglamig

Ku Hundzuka ka Ku Pfuxeta ka Data Center: Ku suka eka Maendlelo lama Sekeriweke eMoyeni ku ya eka ku Pfuxeta ka Mahala

Еволюція охолодження центрів обробки даних: від повітряних методів до вільного охолодження

ڈیٹا سینٹر کولنگ کا ارتقاء: ہوا پر مبنی طریقوں سے مفت کولنگ تک

Ma'lumotlar markazini sovutishning evolyutsiyasi: havoga asoslangan usullardan erkin sovutishgacha

I-Evolution yeZiko leeNkcukacha zokuPholisa: Ukusuka kwiiNdlela eziSekwe kuMoya ukuya kwi-Free Cooling

Ukuvela Kokupholisa Kwesikhungo Sedatha: Kusuka Ezindleleni Ezisekelwe Emoyeni Kuya Ekupholiseni Kwamahhala

Victory was nowhere in sight

Marketplaces are very popular with entrepreneurs today. The advantage of these sites is that a business can immediately start selling goods, and for this, you do not need to create your own website or develop social networks. Popular marketplaces are ready-made online stores where you can register and start trading right away. But just adding photos of goods, their descriptions t, and then sitting and doing nothing will not work. There is a lot of competition on marketplaces, and in order for the buyer to notice your product and not go to another seller, ad texts need to be optimized for key queries - this way your product will be found faster when searching. How to do it correctly and how SEO-optimization of a marketplace differs from a regular site, experts from the agency will tell. Ingenious Guru Search Optimization of The Marketplace: Product Cards and Metadata Search optimization of the product card is necessary for the marketplace search algorithms to see the seller’s offer and show it to potential buyers. SEO for the marketplace works on the same principle as in Google or Bing search engine optimization: the buyer enters a search query and the system offers him products. To get into the TOP of offers, the seller must use keywords in the description of product cards: in the title and characteristics. And you also need to add high-quality photos. What does good marketplace SEO look like to the seller? Product cards rank highly in the internal search results of the marketplace. Potential buyers go to the product card and buy. Customers become loyal, leave positive feedback and come back for repeat purchases. The product card receives additional traffic from Bing and Google search engines. On the other hand, not only does the seller needs optimized product cards, but it is also beneficial for the marketplace. After all, if the seller is engaged in improvement and SEO optimization, then the marketplace as a whole becomes more convenient for customers. And as such, more sellers; all the better for the site. How SEO on the Marketplace differs from Website Promotion When a business is engaged in SEO promotion of its website or online store, it is always a combination of internal and external factors: technical and page optimization; creation and distribution content; link building; crowd marketing; work on awareness and brand reputation; and so on. As a result, we get a huge scope of work. There is also a lot of work in SEO promotion on marketplaces, but the focus is always on one main tool: you need to constantly expand the semantic core. The larger it is, the higher the chance for the seller to get to the TOP of organic search results. A large semantic core is available to sellers with a wide range of products. For example, you sell sports goods in Moscow and you have many different categories: goods for boxing, football, basketball, gymnastics, ballet, dance, martial arts, and so on. For each category, you can have hundreds of products and for all of them, you need to choose keywords and phrases. You select 5-10 keywords for each and add them to the descriptions on the landing pages. If everything is done correctly, the search engine will rank your store highly on the marketplace and display pages in the first positions in the search results. Say your competitor also sells sporting goods on the same marketplace, but he focuses on only two areas: sports uniforms for football and soccer. You can make a large semantic core here, and, most likely, it will be more difficult for him to get into the TOP than you. How to Carry Out Search Optimization of Product Cards on The Marketplace and Reach the top Here are our recommendations to the seller that will help you get to the TOP of the search and which you can do yourself: Use keywords from semantic core in titles to product descriptions. The title should be no longer than 100 characters and it can fit up to six keywords. Make product descriptions detailed to cover all the semantics so that the user does not have any questions. Specifications should be easy to read. At the same time, you also need to add keywords and phrases to them: this helps to bring products to the TOP of the search. Product descriptions should not consist of keywords alone. Because marketplace algorithms can consider such a card as spam or even impose sanctions on the seller. Description text must be unique and not repeated. Specify headings and subcategories in the product cards. And then your products will be shown to users from neighboring categories. Add more quality photos to product cards. From different angles and in good light. The better the photo quality, the more attractive your product looks, the higher the chance that it will be bought faster than that of the competition. Shoot short video reviews of products. This will increase the confidence of potential customers: they will be able to see in advance what their purchase will look like. And they will make a decision in your favor faster. Entrust SEO-optimization of the marketplace to specialists who know all the nuances. Ingenious Guru Agency provides such a service and guarantees a leading position in the marketplace no later than three months of active work. Bonus: What else Affects The Position of a Product on The Marketplace - The Secrets of SEO for The Marketplace We conducted an experiment and found out what else affects the positions of the TOP products in the search for marketplaces: The marketplace displays in the TOP of the results those products that are most often bought and that users of the site are most interested in. The same applies to products that are in stock: the marketplace also displays them in the TOP. Therefore, it is important to monitor the availability and be sure to indicate this in the product card. The better the rating of the seller and the product, the more often it will be shown to users - take care of the reputation and good ratings of the goods and your profile as a seller. It is also affected by the percentage of redemption of the goods: the more often the goods are taken away after being put in the basket, the higher the rating of the seller. Help here buyback - when goods are ordered and turnover increases. More than 80% of buyers on marketplaces are guided by reviews, and if they see at least one negative one, they leave to study similar products from other sellers - it will help in writing positive reviews. Insufficiently complete or errors in the description of goods reduce the likelihood that the seller will get to the TOP of the issue. Therefore, you need to check the text for grammatical and spelling errors before you publish it. The faster the seller answers the buyer's question, the higher the likelihood that the transaction will take place and the goods will be bought. The buyer is ready to wait for an answer for no more than a day. Marketplace SEO Step by Step Various businesses apply to our agency for promotion on marketplaces. We work with sites like Wildberries, OZON, Lamoda, Aliexpress, and others. Our work consists of the following steps that you can take into service: STAGE 1. We Formulate the Problem We do it together with the client. We answer the question “How many sales from the marketplace do you need to get?”. STAGE 2. We Study the Competitive Advantage of the Client's Products It is important to understand what can be emphasized in the description in order to catch the attention of users. STAGE 3. Analyze Everything If the client has already placed his products on the marketplace, we look at the available product cards: what are their characteristics, photos, and whether everything is correct from the point of view of SEO. We compare with the cards of competitors' products in ours, how they are better or worse? STAGE 4. We Create a Promotion Strategy We form the same semantic core, which was discussed above. According to the keywords that we collect, we will promote products. We also include the buyback volume and the number of reviews that need to be placed in the strategy. STAGE 5. Optimizing Product Cards We correct titles and descriptions, order all the texts on the product pages. We develop, if necessary, card designs and for product presentations. infographics STAGE 6. Working on the Seller's Reputation on The Marketplace This is also a very important stage of search engine optimization for the marketplace. And it is based on creating positive reviews about the goods and the seller and posting them on the marketplace, review sites, and social networks. Reviews can be in the format of texts and videos. The more reviews, the higher the seller's rating. STAGE 7. We Launch Advertising We promote product cards on Bing and Google marketplaces using SEO methods. Additionally, on the pages of marketplaces, we provide in Google. Direct, targeted advertising in social networks, and so on. This allows you to increase sales and card conversion. contextual advertising

SEO Promotion for Marketplaces: Optimization of Product Cards 

Hello everyone! My name is Vlad Vinogradov, I am the head of the computer vision department at EORA.ai. We have been engaged in deep learning for over three years and during that time we have implemented many projects for both Russian and international clients, which included the research and training of models. Recently, we have focused on solving the problems of finding similar images and at the moment have created systems for finding logos, drawings, furniture, clothing, and other goods. This post is for Machine Learning Engineers and is based on my talk Finding Similar Images - . A Guide from A to Z, which was published by the Open Data Science community at Data Fest Online 2020 This article provides background information on the recommended methods used in the Image Recovery task. After reading the article, you will be able to create a search engine for similar images for your objective from scratch (not counting the process of developing a production solution). About the project A Similar Image Retrieval (aka Content-Based Image Retrieval or CBIR) is any search that involves images. The picture above from the article will tell you about the problem most easily. Recent Advance in Content-based Image Retrieval: A Literature Survey (2017) Nowadays, the "Search by photo" approach is being used actively, in particular, in e-commerce services (AliExpress, Wildberries, etc.). "Keyword search" (with an understanding of the content of images) has long settled in the search engines Google, Yandex, etc., but has not yet reached marketplaces and other private search engines. I think that since the appearance of the sensational in computer vision circles, the globalization of this approach will accelerate. CLIP: Connecting Text and Images Since our team specializes in neural networks in computer vision, in this article I will focus only on the Search by Photo approach. Basic service components The model can be made on the classic CV or on the basis of a neural network. Model input - image, output - D-dimensional descriptor/embedding. In the case of the classics, it can be a combination of SIFT-descriptor + Bag of Visual Words. In the case of a neural network, a standard backbone like ResNet, EfficientNet, etc. + intricate pooling layers + clever learning techniques, which we will talk about later. I can say that if you have enough data or good practice, neural networks will almost always benefit greatly (we checked), so we will focus on them. Step 1. Train the model. Indexing is a running of the trained model on all images and writing embeddings to a special index for quick search. Step 2. Indexing the base of images. Using the image uploaded by the user, the model is run, the embedding is obtained and this embedding is compared with the rest in the database. The search result is the search results sorted by relevance. Step 3. Search. Neural Networks and Metric Learning The neural network in the task of finding similarities is used as a feature extractor (backbone). The choice of backbone depends on the volume and complexity of the data - you can consider everything from to . ResNet18 Visual Transformer The first feature of the models in Image Retrieval is the magic in the head of the neural network. On the__ __, they are fighting for building the best descriptors - there are Combined Global Descriptors with parallel pools and Batch Drop Block for a more even distribution of activation over the output feature map. Image Retrieval leaderboard The second main feature is the loss functions. There are a lot of them. In alone, there are over a dozen recommended pair losses. There are also the same numbers of classification ones. The main point of all these losses is to train the neural network to transform the image into a vector of linearly separable space, so that further these vectors can be compared by cosine or Euclidean distance: similar images will have close embeddings, dissimilar images will have distant ones. Let's take a closer look. Deep Image Retrieval: A Survey Loss Functions Contrastive Loss This is a double loss, i.e. objects are compared by the distance between each other. The neural network is penalized for the distance from each other of the embeddings of images p and q if these images are actually similar. Similarly, there is a penalty for the proximity of embeddings, the images of which are actually different from each other. In this case, in the latter case, we set the boundary m (for example, 0.5), overcoming which, we believe that the neural network has coped with the task of "separating" dissimilar images. Triplet Loss Here we aim to minimize the distance from the anchor to the positive and maximize the distance from the anchor to the negative. Triplet Loss was first introduced in Google's on facial recognition and has long been a state-of-the-art solution. FaceNet article N-tupled Loss N-tupled Loss- a development of Triplet Loss, which also takes anchor and positive, but uses multiple negatives instead of one negative. Angular Additive Margin (ArcFace) The problem of paired losses is in choosing combinations of positives, negatives, and anchors - if they are just taken uniformly random from the dataset, then the problem of "light pairs" arises. These are such simple pairs of images for which the loss will be 0. It turns out that the network converges quickly enough to a state in which most of the elements in the batch will be "easy" for it, and the loss for them will turn out to be zero - the network will stop learning. To avoid this problem, they began to come up with sophisticated mining techniques for pairs - hard negative and hard positive mining. You can read more about the problem in . There is also a , which implements many mining methods, and in general, the library contains a lot of useful information on the Metric Learning task on PyTorch. this article PML library Another solution to the problem is classification losses. Consider one popular bug feature that led to state-of-the-art facial recognition three years ago - . ArcFace The main idea is to add an indent m to the usual cross-entropy, which distributes the embeddings of images of one class in the region of the centroid of this class so that they are all separated from the clusters of embeddings of other classes by at least an angle m. This seems to be the perfect loss feature, especially when you look at the . But you need to keep in mind that it will work only if there is a classification markup. If you don't have one, you will have to work with paired losses. MegaFace benchmark Here I visually show you which loss functions are best used when you have single-class and multi-class markup (you can derive paired markup from the latter by counting the percentage of intersection between multilabel vectors of examples). Pooling Let's go back to the neural network architecture and consider a couple of pooling layers used in Image Retrieval tasks. R-MAC is a pooling layer that accepts the output map of the neural network (before the global pooling or classification layers) and returns a descriptor vector calculated as the sum of activations in different windows of the output map. Here, the activation of the window is to take the maximum for this window for each channel independently. Regional Maximum Activation of Convolutions (R-MAC) The resulting descriptor takes into account the local features of the image at different scales, thereby creating a rich feature description. This descriptor itself can be an embedding, so it can be sent directly to the loss function. Generalized Mean Generalized Mean (GeM) is a simple pooling that can improve the quality of the output descriptor. The bottom line is that the classical average pooling can be generalized to the lambda norm. By increasing the lambda, we make the network focus on significant parts of the image, which can be important in certain tasks. Ranging Index The key to a high-quality search for similar images is ranking, i.e. displays the most relevant examples for a given query. It is characterized by the speed of building the descriptor index, the speed of the search, and the memory consumed. The simplest thing is to keep the embeddings "head-on" and do a brute-force search on them, for example, using the cosine distance. Problems appear when there are a lot of embeddings - millions, tens of millions, or even more. The search speed is significantly reduced, the amount of heap occupied increases. One positive thing remains - the search quality is perfect with the existing embeddings. These problems can be solved at the expense of quality - to store embeddings, not in their original form, but compressed (quantized). And also to change the search strategy - not to search for brute-force, but to try for the minimum number of comparisons to find the required number of those closest to the given query. There are a large number of efficient frameworks for approximate search for the nearest ones. A special has been created for them, where you can see how each library behaves on different datasets. benchmark Most Popular: , , , . Also, if you want to take indexing with the REST API out of the box, you can consider the . NMSLIB Spotify Annoy Facebook Faiss Google Scann Jina application Re-ranking Researchers in the Information Retrieval field have long understood that ordered search results can be improved by some way of reordering items after the original search results have been received. One such method is Query Expansion. The idea is to use the top-k of the closest elements to generate a new embedding. In the simplest case, you can take the averaged vector, as shown in the picture above. You can also weight embeddings, for example, by distance in the issue or cosine distance from the request. Such improvements are described in a single framework in the article . Optionally, you can apply Query Expansion recursively. Attention-Based Query Expansion Learning k-reciprocal k-reciprocal - a set of elements from top-k, including the k nearest of which is the request itself. On the basis of this set, the process of re-ranking the results is built, one of which is described in the article . By definition, k-reciprocal is closer to the query than k-nearest neighbors. Accordingly, one can roughly consider the elements that are in the k-reciprocal set as knowingly positive and change the weighting rule, for example, for Query Expansion. In this article, a mechanism has been developed for recalculating distances using k-reciprocal sets of the elements themselves in top-k. The article contains a lot of calculations, this is beyond the scope of this post, so I suggest the readers familiarize themselves with it. Re-ranking Person Re-identification with k-reciprocal Encoding Validation We come to the part of checking the quality of the search for similar ones. There are many subtleties in this task that may not be noticed by beginners when they first start working on an Image Retrieval project. Metrics Let's consider such popular metrics in the Image Retrieval task: precision@k, recall@k, R-precision, mAP и nDCG. precision@R Pros Shows the percentage of top-k responses that are relevant. Cons very sensitive to the number of relevant examples for a given query, which does not allow an objective assessment of the quality of search, where there are different numbers of relevant examples for different queries reaching the value 1 is possible only if the number of relevant >= k for all queries R-precision Same as precision@k, where k is set equal to the number of relevant queries. Pros sensitivity to the number k in precision@k disappears, the metric becomes stable Cons you have to know the total number of relevant ones to the request (it can be a problem if not all relevant ones are marked up) Recall@k Shows what proportion of relevant items were found in top-k Pros answers the question of whether found relevant in principle among the top-k stable and well averaged over requests mAP (mean Average Precision) Shows how densely we fill the top of the search results with relevant examples. You can look at this as the amount of information received by the user of the search engine, which has read the least number of pages. Accordingly, the greater the amount of information to the number of pages read, the higher the metric. Pros objective stable assessment of search quality is a one-digit representation of the precision-recall curve, which itself is rich in information for analysis Cons you have to know the total number of relevant ones to the request (it can be a problem if not all relevant ones are marked up) You can read more about metrics in Information Retrieval, including the mAP output, here nDCG (Normalized Discounted Gain) This metric shows how correctly the elements in top-k are ordered among themselves. We will not consider the pros and cons of this metric, since this is the only metric in our list that takes into account the order of elements. However, there are studies showing that this metric is fairly stable when it is necessary to take into account the order and can be suitable in most cases. Estimation a metric is calculated for each request, averaging over all requests macro: pros: there are no significant fluctuations with a different number of relevant to this query cons: all queries are considered equal, even if some are more relevant than others the number of tagged relevant and separately successfully found relevant is summed over all queries, then participates in the fraction of the corresponding metric micro: pros: queries are evaluated taking into account the number of tagged relevant for each of them cons: the metric can become very low / very high if there were a lot of tagged relevant ones for some request and the system unsuccessfully / successfully brought them to the top Validation system proposing to consider two options for validation Validation on a set of queries and selected relevant ones Input: image requests and images relevant to them. There is markup in the form of a list relevant to this query. To calculate metrics, you can calculate the relevance matrix each with each and, based on the binary information about the relevance of the elements, calculate the metrics. Full-base validation Input: image requests, and images relevant to them. There should also be a validation database of images, in which, ideally, all relevant queries are marked. Also, it should not contain query images, otherwise, they will have to be cleaned at the search stage so that they do not clog the top-1. The validation base is involved as a base of negatives - our task is to draw out those that are relevant to it. To calculate metrics, you can go through all the requests, calculate the distances to all elements, including relevant ones, and send them to the metrics calculation function. Completed project example Some companies argue with other companies that the latter should not use the pictorial elements of the former's brand. In such cases, a weaker manufacturer tries to parasitize on the success of a large brand by presenting its products and services under similar symbols. Users also suffer from this - you can mistakenly buy cheese from the wrong manufacturer, which you already trust, but take a fake product without carefully reading the label. An example from a recent one: Apple Is Attempting to Block a Pear Logo Trademark There are special companies, governmental and private, to fight illegal immigration. They have a register of registered trademarks by which they can compare new incoming trademarks and approve / reject an application for the registration of a trademark. Above is an example of the interface of the system. In such systems, a search for similar images will be a good helper - the expert will be able to find analogs faster. WIPO Examples of how our system works The size of the indexed database of images: several million trademarks. Here the first image is a query, on the next line is a list of expected relevant ones, the rest of the lines are what the search engine gives out in order of decreasing relevance. Conclusion That's all. It was overview material. Hopefully, those who are building or about to build similar image search engines have gained some benefit. And vice versa, if you think that I am wrong somewhere, tell me in the comments, I will be glad to receive feedback.

How to Build an Image Search Engine to Find Similar Images

Stories, News & Mentions about WILDBERRIES

Wildberries’s Stories on HackerNoon

Wildberries's News & Mentions