Uber Eats is a large-scale marketplace that connects users and restaurants around the world. The speaker explains that the service operates at a run rate of about 90 billion, adds millions of new items each month, is growing about 20% year on year, and spans 10,000 cities. The unit and currency for the roughly 90 billion are not given in the talk.
As an additional explanation, at this scale, a manual process for only some dishes, or a one-off system designed for one region, is difficult to use. The service needs an automated and general approach that can handle images from many restaurants, dishes, and places. Processing must also be fast and safe for users, and the team must be able to check what happened afterward. Feedback is also needed so the team can see results in operation and improve the system.
In other words, the image problem is not just image processing for a demo. It is a real service problem that must handle a large and continuously growing set of new items. How images are handled is also connected to the experience of the marketplace as a whole.
The user path described by the speaker is as follows. First, a food photo creates the user's initial impression of a restaurant. The user scrolls through the feed, clicks an item that looks interesting, and finally adds it to the cart.
Example (additional explanation): Even for the same dish, a clear photo may make a user more likely to open the item page to check its contents. From there, the user may continue to adding it to the cart. However, the information in this chapter does not tell us by what percentage a photo changes the add-to-cart rate.
This path does not mean that a photo always causes a purchase. The speaker's point is that the photo is at the beginning of a path that can lead to user action. Image quality therefore concerns more than appearance; it is related to the user experience and behavior in the marketplace.
モダリティが増えると、評価する対象も広がる
As modalities increase, so does the evaluation surface
A modality is a type or format of information. Here, a food photo is one modality: a still image. The speaker says that video, in addition to photos, is becoming increasingly important.
With only still images, the system can focus mainly on composition, color, and how the food appears. With video, changes over time also become part of the input. When formats such as photos and video increase, the range that the system must understand, process, and check for quality also becomes wider. The speaker does not state video's share or which other modalities are included.
The need, then, is not simply a system that makes one image look better. It is a system that handles varied visual content reliably within an experience used by people around the world. The eval must check not only visual appeal, but also whether the system fits the input types and the user experience.
The slide places a feed of restaurant and dish thumbnails beside a vertical food video on a smartphone, under a heading about visual content and the user experience. It lets us see the context behind the speaker's explanation that food images and video are an entry point to the service. The small text inside the smartphone is too small to read accurately.
At the scale of a marketplace such as Uber Eats, the system must handle a diverse set of items that grows every month, along with content expanding from photos to video. Food images are the entry point to an experience in which users learn about a restaurant and item and move from the feed to the cart. An image-quality system therefore has to consider scale, varied input formats, and the user experience at the same time, rather than merely making images look polished.