Core demands
With the gradual improvement of consumption level, home decoration has evolved from meeting the basic living needs to meeting the needs of a better life. At present, users' spending on home decoration is also growing at an annual rate of 10%. Therefore, personalized home has become an important consumer demand. For Home Map, how to let users find their favorite pictures in the platform, and then find the products in the pictures they like to purchase is an important embodiment of personalization. Because the home design pictures involve certain professional knowledge of home design, the traditional way is that the platform operators manually classify and label them, which is inefficient. However, the accuracy rate is very low when uploaders are allowed to tag by category through the platform. The low efficiency and low accuracy of the basic classification of pictures make it impossible for users to quickly improve the conversion rate of goods purchase.
Solution
Customized image recognition solution: allow customers to customize their own image recognition models, and complete model training only by labeling a small amount of data. The advantages of this scheme are:
1. Submit training pictures by dragging and pulling, and quickly complete data annotation and model training;
2. A variety of algorithm components and training templates, based on Baidu big data to achieve a small amount of data training precision model;
3. Provide one-stop service of data annotation model training stable API generation.
The traditional method is that the demander submits a dataset, the technical service provider manually creates a service, and then delivers the API to the demander after training. This method is inefficient. If the demander wants to train a large number of classification labels at the same time, it not only requires a large amount of data for training, but also has a long cycle. Using Baidu's customized image recognition solution, we can open multiple training sets at the same time to classify and label household pictures in multiple dimensions.
Let's take the hybrid application of the "space, color, style" classification tag we trained as an example to illustrate the application scenario of customized image recognition.
Application Scenario 1: Guess you like it (recommended for similar pictures)
Our requirements for "Guess what you like" are:
It is necessary to recommend the pictures in the same space. For example, in the above figure, if the current picture is a picture of the living room, then the "Guess what you like" below must be a picture of the living room;
Must be close in style and color to ensure visual similarity.
Only when the above two conditions are met at the same time can users really find the pictures that meet their interests. We used Baidu's "customized image recognition platform" to train classification tags in three dimensions of space, color and style in a very short time, and then obtained the current results through cross algorithm. Data proves that the new "Guess what you like" has increased the number of users' image clicks by 30%.
Application Scenario II: Product identification and purchase recommendation in the picture (photo identification)
Similar problems can be solved through image recognition technology, but in the home picture scene, many goods will have different uses due to different placement spaces, and the recommended goods should also be different. For example, "cabinet", a cabinet with similar shape, if it is placed in the porch, the most likely is "shoe cabinet"; In the dining room, it should be the "sideboard". This is not a technical category, but an industry knowledge category.
In order to solve the above problems, we use Baidu's "customized image recognition technology" to label each image with a space tag, and then cross it with the identified commodity category to get the correct results.