IBM Watson Visual Recognition alternatives after its shutdown

IBM Watson Visual Recognition is discontinued. Support ended on December 1, 2021, and IBM stopped allowing new instances on January 7, 2021. IBM's SDK documentation also states that remaining instances would be deleted. An old tutorial or SDK reference does not mean the cloud service is still available.

If you are maintaining an application that used it, choose a replacement for the job the application performed. General image labels and a classifier trained to recognize your own products need different evaluations. Start with the images, labels and application code you still have, rather than planning around access to the retired service.

Match the replacement to the old workflow

These are documentation-based candidates, not tested equivalents or a ranking of accuracy. They also require changes to integration and evaluation; do not assume an old Watson classifier can be uploaded into another service.

General labels do not replace your custom categories automatically

Google Cloud Vision label detection identifies general entities such as objects, locations and activities. Its response includes labels and confidence scores. It is a reasonable candidate when your application needs broad image tagging.

Suppose your old application separated your own catalog into specific product families. A service that recognizes a chair may still fail to distinguish the categories your business uses. Compare the proposed labels with the application's expected outputs before changing the API call. Decide how the application should handle unfamiliar labels and uncertain results.

A custom classifier needs training material you can still use

Amazon Rekognition Custom Labels supports image-level classification and locating custom objects within images. Its workflow uses labeled training images and provides evaluation metrics on a test set.

Check what remains from the old project: original images, category definitions, annotations and examples reserved for evaluation. If those assets are missing, budget for collecting or labeling a new dataset. Retain a test set that represents difficult cases as well as the easy examples.

Custom Labels is not designed for text detection, face analysis or unsafe-content detection. AWS identifies other Rekognition features for those tasks. Choose by the specific capability rather than treating everything under a product name as interchangeable.

Keep text extraction separate from classification

For text in photos, Cloud Vision OCR offers TEXT_DETECTION. DOCUMENT_TEXT_DETECTION is optimized for dense text and documents, with a response organized into elements such as pages, paragraphs and words.

A workflow that sorts invoices and extracts their contents may need both classification and text processing. Test each part separately. Correctly recognizing an invoice does not establish that its totals, dates or line items were extracted correctly.

Rebuild the contract between the model and the application

Before replacing the old integration, list the outputs the application depends on: label names, object locations, confidence thresholds and error responses. Map the new service's response into that contract, and test what happens when no result meets your threshold. Do not carry a threshold across services without checking its effect on your own examples.

Run representative images through the candidate, review the errors and measure the complete request time. For a custom model, include training and model-running costs in the estimate; for a pretrained API, check the billing rules for each requested feature.

AWS bills Custom Labels for training and inference hours. An available custom model can keep incurring charges even when it processes no images. If you process a daily batch, include the work of starting and stopping the model in your plan rather than estimating cost from image count alone.

Use retained assets you have permission to process. Old SDK export methods are historical documentation, not a recovery path for deleted IBM instances. Once the replacement meets your application's requirements, test the surrounding workflow before removing the old integration code.