Skip to content

Latest commit

 

History

History
41 lines (27 loc) · 3.09 KB

CustomVisionExplorer.md

File metadata and controls

41 lines (27 loc) · 3.09 KB

Custom Vision Explorer

This example shows how to use the Custom Vision APIs and how to score images against a custom image classifier or object detector.

Once you have it configured and at least one project created, just select a target project and send an image for scoring. You can do so from local files, Bing Image Search results or from a web camera capture. Once you submit a photo, the results are shown on above the photo as a list of tags and their associated confidence level.

Image Classification model: alt text

Object Detection model: alt text

Active Learning

It is very easy to create and train a project to do image classification on popular subjects (literally a few seconds as it integrates nicely with Bing Image searches). Scoring against these models using Bing Images works well, but if you use a web camera capture for scoring it is easy to run into issues. Between the other stuff around an object, or the unique lighting and POV from web camera captures, things can be very different than the images used in the trained set.

As such, a good way to enhance a classifier is by correcting any missed result for images used during scoring. When you score an image against one of your projects, a message will be displayed above the results (red arrow below) that allows you to use that image, along with the corrected result that you provide, as an input signal back into your project. All you do is provide the proper tags and the kiosk does the rest – it will add the image to the proper tag(s) in your project and trigger a re-train of the project.

Here is the UI that pops up when you enter the Active Learning UI. Just toggle all the tags that should be associated with the image and click the big “Save and re-train the project” button.

alt text

As of now you can only use this workflow to correct images that still belong to at least one of the tags in the project. If you need to correct images that shouldn't belong to any tags, the current approach is to create an Other tag an associate such images to that tag.

Key Source Code

  • CustomVisionExplorer: Main page that drives the demo. It hosts the CameraControl (see below) to display the live camera feed and allow photo captures, as well as the ImageSearchUserControl as another input model.

  • CameraControl: The code that contains the camera feed and enables camera captures.

  • ImageSearchUserControl: A control used for providing input images. It allows you to provide images from Bing Images, local images or the web camera.