Image Search @ Tokopedia.com

Just imagine, you go to a friend’s birthday bash and you get mesmerised by an evening gown or a tuxedo that one of the guests is wearing.

You take a snap of the person by some excuse and scour through the e-commerce portals and retail shops looking for exactly the same dress. You will probably spend a couple of hours and if you are very lucky you will find the right match otherwise you will give up. Most of the time we give up and life moves on.
Been there, done that, right ?

Well this is exactly the problem we tried to solve at Tokopedia for our customers. We simplified this process to three steps:

  • Open the Tokopedia App, click on the search bar
  • Click on camera(icon) for taking a photo.
  • Snap a photo of the product, you want to search. Click Ok
I am looking sad because Argentina got kicked out from WC 2018

This seems to be a magic right ? Go ahead and try for yourself to experience the magic.

If you are interested in a video, you can see here: https://youtu.be/UsRYeFfiSaY

If you have a downloaded an image from Instagram/FB, follow this video: https://www.youtube.com/watch?v=GfUQQf5r590

Enter the Dragon

We at Tokopedia were working on our own image recognition and image search services when Alibaba offered publicly available Image Search Service on the Aliyun Cloud. Alibaba is using the Image Search extensively on the Taobao portal as the Pailitao (snap and buy) feature. So we hopped onto the offer and a couple of POCs later decided to use the service in Tokopedia.

Image Search is an intelligent image service based on deep learning and machine learning technology. By incorporating image recognition and search, it allows you to search by image across different industries and business scenarios. Image Search is developed to meet the requirements of searching and finding similar or identical images.

According to the Alibaba documentation, Image Search provides the following benefits:

High accuracy: provides industry-leading search services.

Fast response: responds to queries in milliseconds.

Large volume support: supports over ten billion items.

Real-time interface: allows you to add or remove data dynamically.

Customization: provides flexibility for your individual requirements.

Our Rating:

High accuracy: +1, the results are quite accurate

Fast response: -1. Not so fast, we found that the average time was 0.9 to 1s.

Large volume support: +1, we uploaded 30 million product images.

Real-time interface: +1, we can update data dynamically.

Customization: -1, we did not find much flexibility for custom requirements.

Throughput: -1, current throughput is 7 qps which is less.

Given the fact that Alibaba Image search is still in nascent stage and since the accuracy was good, we got a go ahead to launch the feature for a limited number of users on Android. Encouraged by the response we got, we made it available to 100% users in the month of May.