The search service is a key feature of ProductAI. It helps customers to build index on their own image set, to achieve a series of functions such as photo&buy, product management, product association, etc. Currently, the search service of ProductAI have following features:
- ProductAI not only supports image-based search, but also supports price filtration and fuzzy word search, see Search Type
- Customize search models for different industry application scenarios, see Scenarios
- Support non-real-time indexing. Easy to use, and fast to upload image data, see Indexing Type
ProductAI supports two types of index and search service:
- Image Search: The engine will index massive images and use image features to find similar items of the query image.
- Product Search: The engine will index massive product information, and use product images, keywords in the title and the product price to query and filter products.
- Real-time Indexing: index data while uploading. Once a data item has been uploaded successfully, we can query the item immediately. Uploading data will be a time-consuming task since uploading and indexing are synchronized processes. Real-time indexing is suitable for the situation that the amount of data is small and the data need to be indexed in real-time.
- Non-real-time indexing: index data by an asynchronized system. Users can query this search service right after this service is created. But please note that at this initial stage only a few images can be fetched because the indexing process is just started. The whole indexing stage will finish after a certain time frame. The total time cost depends on image data size, scenario model complexity and infrastructure resources used on this service which is scheduled by ProductAI. Please contact our account manager if you need a priority indexing service. This indexing type is suitable for the situation that the amount of data is massive but the data have no need to be indexed in real-time.
Notes: ProductAI only supports non-real-time indexing currentlty.
Image search service uses different model for different scenario, for example: you can expect a model trained on furnitures to gain a good performance if you use this model on furniture scenario, but it is highly possible to see a bad performance in a fashion clothes scenario. To analogy with traditional relational database, a scenario can be recognized as a index of data table. Supported scenarios of ProductAI are listed below:
|Scenario name||Scenario ID(Used as parameter by SDK/API to initiate scenario)||Current version||Description|
|Fashion||fashion_5_7||v5.7||For fashion image search scenario which uses PGC (Professional Generated Content) and UGC data as traning set, it fits in fashion image search product management systems. Reference: 19 Fashion Tags|
|Shoe||shoe_v1||v1.0||For shoe image search scenario which uses PGC (Professional Generated Content) and UGC data as training set. Suitable for shoe e-commerce search products and product management systems|
|Furniture||furniture_v6||v6.0||For furniture image search scenario which uses PGC (Professional Generated Content) and UGC data as traning set, it fits in furniture image search product management systems|
|Wine||wine_v2_1||v2.1||For wine image search scenario which uses UGC data as traning set, it fits in wine image search product management systems|
|Material||material_v2||v2.0||For material image search scenario which uses PGC (Professional Generated Content) and UGC data as traning set, it fits in material(especially on patterns) image search product management systems. For best results please use flat material images when creating dataset. Our models are optimized for flat material images. See Reference 2|
|General||general||v1.0||For general image search scenario which uses PGC (Professional Generated Content) data as traning set, it fits in general image search systems|
|Ecommerce Product||ecom_product_2_1||v2.1||For e-commerce product image search scenario which uses PGC (Professional Generated Content) data as traning set, it can be used to search similar product|
|Precise Creative Picture Matching||dedupe_tuku_v2||v2||For precise creative picture matching scenario, see Precise Creative Picture Matching|
|Precise E-Commerce Image Matching||dedupe_v2||v2||For precise E-Commerce image matching scenario, see Precise E-Commerce Image Matching|
Reference 3: 19 Fashion Tags
Accurate matching is a special image search scenario. It filters and re-orders the images after the normal image search model, in order to return the most accurate matching result with query image. It supports two scenarios currently :Precise Creative Picture Matching and Precise E-Commerce Image Matching.
Since matching service is also a use case of image search, all the operations of search service can be used on matching service
This service builds index for creative photos, art designs, paintings and movie pictures. User is able to find images that exactly the same or with a certain range of modifications (such as waterprint, add image effect filter, etc) compared to query image. This service can be used in IP protection of creative images or products, video content search, etc.
|Query Image||Search Result: water print, rotation, clipping, scaling, mirroring, color space shifting, etc|
This service indexes product images, user is able to find images that exactly the same or with a certain range of modifications (such as waterprint, add image effect filter, etc) compared to query image. This service can be used in product price comparing, relative searching, de-duplication, etc.
|Query Image||Search Result (even with some modifications, such as water print)|
|Query Image||Search result, with some pose difference|