Reviewing Tag Results for your Items with VueTag

Introduction

VueTag allows users to upload images as batches, or send images via the API to obtain generated tags. These predictions (resulting tags) are displayed in your Vuetag account or made available in the API response. This article talks about various scenarios on how the tags can be reviewed and exported.

It's important to identify what you'd like to review for, based on your business needs. For example:

  1. The top level category predictions (For example: Pants, Dresses etc ) for the images are likely to be picked first for review since you may need to ensure that this top level product type is identified correctly.
  2. Your business need could be to standardize an attribute (for example: colors ) of the products in your catalog. In this case, you may choose to review (for example) the Base Color predictions according to your needs.

Another way to determine which images to review could be on the basis of volume of images.

  1. If the volume of images is low and users would like to use all the tags predicted by VueTag, this can be done using any of the methods described below
  2. If the volume of images is high, users can review a subset of the images under each category for the predictions to get an idea of the accuracy of the VueTag systems

Note: Review of tags includes accepting the tag predictions generated by the system and editing the incorrect or no predictions to the correct value. The accuracy of the category or attribute is calculated on the reviewed images.

Reviewing images

1. Bulk Edit:

  • When the attributes to be reviewed are determined, users can use Bulk Edit to provide review for a set of images in one go. Click on the Bulk Edit icon on the category box to go to the Bulk Edit screen

  • By default, the bulk edit page loads with the Category filter applied

  • Users can see the predicted values on the images at the bottom.
  • Users can select the images where there are incorrect predictions and "move" them to the correct value on the right side pane.
  • The filters on the left panel for the attributes can also be chosen to filter images by the predictions for that attribute.

  • Sort options: There are three sort options provided on the Bulk Edit page. By descending order of confidence score, ascending order of confidence score and a random order. By default, the images are sorted randomly to have an unbiased set of images to review.
  • The Pending and Reviewed tabs show images pending reviews and what has been completely reviewed. Under the Reviewed section, users can see those images which have been reviewed by them or by other team members.

For more information on the Bulk Edit feature, read here
 

Note: Support for no predictions and multi value attributes is coming soon in Bulk Edit!

2. Single image edit:

If the volume of images tagged is less and your business would like to use all the tags predicted by VueTag, editing via the single image mode is recommended.

  • Click on the category box to go into a Category
  • Click on the image in the center pane to open the image detail page. Here all the tag predictions for that image are displayed.
  • Based on the type of image and/or the confidence of the networks to identify certain attributes, the prediction will be displayed in the drop down.
  • The confidence score of the network is also displayed on the far right column, indicating how certain the system considers the predicted value.
  • Users can change the prediction by choosing another value from the drop down on the attribute which needs editing/
  • Once done, the user can save the image and go to the next image. After which, the other attributes which have not been edited are considered as accepted.
  • Multi value attributes like Style and Occasion can have multiple values when review is provided.
  • Attributes for which users would like to remove the predicted value - can be done by unchecking the prediction in the drop down.

Note: All review which is provided via the VueTag dashboard is routed back to VueTag's internal systems. This review is supervised and used for re-training the networks.

Accuracy report

Once review is done, users can see the accuracy (of prediction) for the reviewed categort/attribute in the Accuracy report for a batch. If category level review is completed under a batch, the confusion matrix shows the comparison of the Prediction Vs. Actual (based on the feedback provided). You can read more about the accuracy report here.

Exporting reviewed images

By exporting the tags from VueTag, users can export the reviewed tags in a CSV file and use this downstream in their business process - like importing it into their PIM system or any other in house content management system. The fetch_reviewed and fetch_tag APIs are also available to download the images which have been reviewed and Once the tags are exported, users will see the final reviewed values for the category and attribute(s) which were reviewed in the dashboard.

  • 96
  • 09-Aug-2019
  • 122 Views