What's new

The following functions, features, and support have been added for PowerAI Vision Version 1.1.3:

PowerAI Vision Non-production edition

You can now try PowerAI Vision for one year with the Non-production edition. This edition does not allow you to export data sets or models.

GPU sharing for deployed models

The full version of PowerAI Vision now supports GPU sharing for deployed models. Deploying multiple models to a single GPU allows you to get the most out of your processing power. GPU sharing is supported only for GoogleNet and Faster R-CNN models. For more information, see Deploying a trained model.

Train with a Detectron model

You can now use a Detectron model to train a model. This allows you to train with objects that have been labeled as non-rectangular shapes. For details, see Training a model.

Transfer learning

You can use a model that was previously trained with PowerAI Vision as a base model to train new models. For details, see Training a model.

Use non-rectangular shapes when labeling

When labeling objects in a data set that will be used to train a Detectron model, you can use non-rectangular shapes. Non-rectangular labeling is supported in images, video frames, and with auto labeling. If you label objects with non-rectangular shapes and train the data set using a different model, associated rectangular bounding boxes are used. For more information, see Labeling objects.

Support of COCO annotations

Images with COCO annotations can be imported. Only object detection annotations are supported. For more information, see Importing images with COCO annotations.

Downloadable heat map

You can download the heat map that is generated when testing an image with a deployed model.

Improved performance for inference

Speeds when using the image classification (GoogLeNet) and object detection (Faster R-CNN) models for inference are improved. The improvement is especially significant for high-resolution images.

Improvements to the user interface

The following changes have been made to the user interface to improve your experience:

• Heat map overlay: When testing an image with a deployed model trained for classification, the heat map is layered on top of the image. You can then use a slider to set the opacity. This allows you to easily identify which areas of the image the algorithm is focusing on.
• Confidence threshold slider: When testing an image with a deployed model, you can use the confidence slider to eliminate object labels that have low confidence.
• GPU information: You can view how many GPUs the system can access and how many of those are in use on the Models or Trained Models page. See Working with the user interface for details.
• Improvements to labeling
  • The working image is given more screen space.
  • New Objects panel consolidates information about labeled objects and has new settings for labeling.
  • Labels on the image are shortened to two characters; with a corresponding list in the Objects panel.
  • You can use standard keyboard shortcuts to copy a shape that you traced and paste it elsewhere in the same image or to any image in the image carousel.
  • You can undo and redo shape creation, edits, and deletions via standard keyboard shortcuts.
  • A Paste previous button was added when labeling videos. Clicking Paste previous copies all the labels from the previous video frame and paste them into the current frame.
  • New settings let you customize your labeling process. For example, you can change the outline color, hide previously drawn outlines, show or hide labels, and so on.
  • Keyboard shortcuts have been added to speed up image navigation and enhance shape management.
  • The list of labeled objects can be filtered.