This topic answers frequently asked questions about Image Search.
Differences between product and generic image search
Product Image Search is designed for product images. It classifies images into relevant categories and extracts fine-grained features for each category to ensure high recall rates. This makes it ideal for e-commerce applications. Generic Image Search does not require category information. All images are treated as part of a single, general category. When you perform a search, you do not need to specify a category for the image because the system handles it automatically. This is suitable for applications like image galleries and web forums.
Reasons for buildPostContent call failure
This error usually occurs for one of two reasons: a required field is missing, or the request body exceeds 8 MB.
Cross-region OSS buckets for incremental updates
No, OSS does not support cross-region data access. If your instance is in the Asia Pacific SE 1 (Singapore) region, the configured OSS bucket must also be in the Asia Pacific SE 1 (Singapore) region.
Category ID for image queries and additions
For Product Image Search, you can choose not to specify a category ID in the add and search API calls. In this case, the system automatically predicts the category. If you specify a category ID, the system uses the specified ID. For Generic Image Search, you do not need to specify a category ID. Any specified category ID is ignored.
Using the crop and region parameters
If you set the crop parameter to true or do not specify it, the system performs main object detection. In this case, if you specify the region parameter, that specified region is used as the main object. If you do not specify the region parameter, the system automatically detects the main object. If you set the crop parameter to false, the system skips main object detection and uses the entire image for feature extraction.
Understanding SortExprValues
SortExprValues is a semicolon-separated pair of values. The first value is the image relevance score. A higher score indicates greater relevance to the query image. The value ranges from 0 to 7.33136443711219e+24. The score is highest when two images are identical. The second value is the Hamming distance of the image feature vectors. A higher distance indicates lower similarity. Results are sorted by the score, so you only need to consider the first value. There is no specific threshold to determine whether two images are similar.
Reasons for incremental task submission failure
Common causes include the following:
The selected OSS bucket and the Image Search instance are in different regions. Supported regions are Asia Pacific SE 1 (Singapore), China (Hong Kong), Asia Pacific NE 1 (Tokyo), and Europe Central 1 (Frankfurt).
The created role is not an Image Search service role. The authorization policy for the service role was not edited correctly, or the policy does not grant the required permissions to the data directory in the bucket. The provided ARN or bucket is invalid, the path is incorrect, or the
increment.metafile is missing from the specified path.
Renewing expired instances and reusing instance names
You can manually renew an instance within seven days after it expires. After seven days, the instance is released and cannot be renewed. You must create a new one. You can customize the name of a new instance, but the name must not conflict with an existing instance name.
API call limits
Image Search does not limit the total number of calls, but it limits QPS based on your instance specification.
Finding the free trial instance
To find your free trial instance, switch to the region you selected during creation by using the region selector in the upper-left corner of the console.
Request limits for the 100,000-image specification
The number of search requests is not limited, but QPS is limited to 1. If you require higher QPS, upgrade your instance in the console.
The llms.txt file for AI agents
Documentation link: https://www.alibabacloud.com/help/en/image-search/llms.txt.