What's new in Custom Vision

Learn what's new in the service. These items may be release notes, videos, blog posts, and other types of information. Bookmark this page to keep up to date with the service.

May 2022

Estimated Minimum Budget

  • In Custom Vision Portal, users can now view the minimum estimated budget needed to train their project. This estimate (shown in hours) is calculated based on volume of images uploaded by user and domain selected by user.

October 2020

Custom base model

  • Some applications have a large amount of joint training data but need to fine-tune their models separately; this results in better performance for images from different sources with minor differences. In this case, you can train the first model as usual with a large volume of training data. Then call TrainProject in the 3.4 public preview API with CustomBaseModelInfo in the request body to use the first stage trained model as the base model for downstream projects. If the source project and the downstream target project have similar images characteristics, you can expect better performance.

New domain information

  • The domain information returned from GetDomains in the Custom Vision 3.4 public preview API now includes supported exportable platforms, a brief description of model architecture, and the size of the model for compact domains.

Training divergence feedback

  • The Custom Vision 3.4 public preview API now returns TrainingErrorDetails from the GetIteration call. On failed iterations, this reveals whether the failure was caused by training divergence, which can be remedied with more and higher-quality training data.

July 2020

Azure role-based access control

  • Custom Vision supports Azure role-based access control (Azure RBAC), an authorization system for managing individual access to Azure resources. To learn how to manage access to your Custom Vision projects, see Azure role-based access control.

Subset training

  • When training an object detection project, you can optionally train on only a subset of your applied tags. You may want to do this if you haven't applied enough of certain tags yet, but you do have enough of others. Follow the Client library quickstart for C# or Python to learn more.

Azure storage notifications

  • You can integrate your Custom Vision project with an Azure blob storage queue to get push notifications of project training/export activity and backup copies of published models. This feature is useful to avoid continually polling the service for results when long operations are running. Instead, you can integrate the storage queue notifications into your workflow. See the Storage integration guide to learn more.

Copy and move projects

  • You can now copy projects from one Custom Vision account into others. You might want to move a project from a development to production environment, or back up a project to an account in a different Azure region for increased data security. See the Copy and move projects guide to learn more.

September 2019

Suggested tags

  • The Smart Labeler tool on the Custom Vision website generates suggested tags for your training images. This lets you label a large number of images more quickly when training a Custom Vision model. For instructions on how to use this feature, see Suggested tags.

May 2019

  • Bug fixes and backend improvements
  • Improved portal UX experience related to Azure subscriptions, making it easier to select your Azure directories.

April 2019

  • Increased limit on number of bounding boxes per image to 200.
  • Bugfixes, including substantial performance update for models exported to TensorFlow.
  • Added Object Detection export for the Vision AI Dev Kit.
  • UI tweaks, including project search.

March 2019

  • Custom Vision Service has entered General Availability on Azure!
  • Added Advanced Training feature with a new machine learning backend for improved performance, especially on challenging datasets and fine-grained classification. With advanced training, you can specify a compute time budget for training and Custom Vision will experimentally identify the best training and augmentation settings. For quick iterations, you can continue to use the existing fast training.
  • Introduced 3.0 APIs. Announced coming deprecation of pre-3.0 APIs on October 1, 2019. See the documentation quickstarts for examples on how to get started.
  • Replaced "Default Iterations" with Publish/Unpublish in the 3.0 APIs.
  • New model export targets have been added. Dockerfile export has been upgraded to support ARM for Raspberry Pi 3. Export support has been added to the Vision AI Dev Kit.
  • Increased limit of Tags per project to 500 for S0 tier. Increased limit of Images per project to 100,000 for S0 tier.
  • Removed Adult domain. General domain is recommended instead.
  • Announced pricing for General Availability.

February 2019

  • Announced the end of Limited Trial projects (projects not associated with an Azure resource), as Custom Vision nears completion of its move to Azure public preview. Beginning March 25, 2019, the CustomVision.ai site will only support viewing projects associated with an Azure resource, such as the free Custom Vision resource. Through October 1, 2019, you'll still be able to access your existing limited trial projects via the Custom Vision APIs. This will give you time to update API keys for any apps you've written with Custom Vision. After October 1, 2019, any limited trial projects you haven't moved to Azure will be deleted.

January 2019

  • Support added for new Azure regions: West US 2, East US, East US 2, West Europe, North Europe, Southeast Asia, Australia East, Central India, UK South, Japan East, and North Central US. Support continues for South Central US.

December 2018

  • Support export for Object Detection models (introduced Object Detection Compact Domain).
  • Fixed a number of accessibility issues for improved screen reader and keyboard navigation support.
  • UX updates for image viewer and improved object detection tagging experience for faster tagging.
  • Updated base model for Object Detection Domain for better quality object detection.
  • Bug fixes.

November 2018

  • Added support for Logo Domain in Object Detection.

October 2018

  • Object Detection enters paid preview. You can now create Object Detection projects with an Azure resource.
  • Added "Move to Azure" feature to website, to make it easier to upgrade a Limited Trial project to link to an Azure. resource linked project (F0 or S0.) You can find this on the Settings page for your product.
  • Added export to ONNX 1.2, to support Windows 2018 October Update version of Windows ML. Bug fixes, including for ONNX export with special characters.

August 2018

  • Added "Get Started" widget to customvision.ai site to guide users through project training.
  • Further improvements to the machine learning pipeline to benefit multilabel projects (new loss layer).

June 2018

  • UX refresh, focused on ease of use and accessibility.
  • Improvements to the machine learning pipeline to benefit multilabel projects with a large number of tags.
  • Fixed bug in TensorFlow export. Enabled exported model versioning, so iterations can be exported more than once.
  • Bug fixes & backend improvements.
  • Enabled multiclass classification, for projects where images have exactly one label. In Predictions for multiclass mode, probabilities will sum to one (all images are classified among your specified Tags).

May 2018

  • Introduced preview Object Detection feature for Limited Trial projects.
  • Upgrade to 2.0 APIs
  • S0 tier expanded to up to 250 tags and 50,000 images.
  • Significant backend improvements to the machine learning pipeline for image classification projects. Projects trained after April 27, 2018 will benefit from these updates.
  • Added model export to ONNX, for use with Windows ML.
  • Added model export to Dockerfile. This allows you to download the artifacts to build your own Windows or Linux containers, including a DockerFile, TensorFlow model, and service code.
  • For newly trained models exported to TensorFlow in the General (Compact) and Landmark (Compact) Domains, Mean Values are now (0,0,0), for consistency across all projects.

March 2018

  • Entered paid preview and onboarded onto the Azure portal. Projects can now be attached to Azure resources with an F0 (Free) or S0 (Standard) tier. Introduced S0 tier projects, which allow up to 100 tags and 25,000 images.
  • Backend changes to the machine learning pipeline/normalization parameter. This will give customers better control of precision-recall tradeoffs when adjusting the Probability Threshold. As a part of these changes, the default Probability Threshold in the CustomVision.ai portal was set to be 50%.

December 2017

  • Export to Android (TensorFlow) added, in addition to previously released export to iOS (CoreML.) This allows export of a trained compact model to be run offline in an application.
  • Added Retail and Landmark "compact" domains to enable model export for these domains.
  • Released version 1.2 Training API and 1.1 Prediction API. Updated APIs support model export, new Prediction operation that does not save images to "Predictions," and introduced batch operations to the Training API.
  • UX tweaks, including the ability to see which domain was used to train an iteration.
  • Updated C# SDK and sample.

Azure AI services updates

Azure update announcements for Azure AI services