You can use Zia Vision to automatically validate the images that are added to your CRM's records. Using this feature, you can ensure that these images are accurate, consistent, and of the requisite quality. This has multiple benefits, ranging from adhering to guidelines to providing a positive experience for all users of your CRM.
As of now, the Detect and Match and Detect options are available only in the US DC.
Training data
For both kinds of entities(labels and objects), Zia has a built-in Gallery with folders containing sample images. These cover a few standard categories, such as:
- Bicycles, Motorbikes, Food, and Cars for match
- Laptop, Person, Chair, Sofa, and so on for detect
For most use cases, you will need to train Zia using your own images. These are called Custom Images. You will need at least five images for each entity (label or object) that you wish to detect.
In addition to adding images in this manner, you can also:
- Add images that have been suggested for training by reviewers (depending on your rule's configuration).
- Remove images from training data.
You can have a maximum of 300 custom images for a validation rule. This includes the uploaded images, suggested images, and approved images.
Once there has been a change of at least 40% in the training data (including both additions and removals), Zia will retrain herself and start validating images based on the updated training data.
Metrics
Success rate
The success rate measures how accurate Zia's image validation has been. It compares the number of accurately validated images to the total number of validated images. It is calculated periodically by Zia and is reset when the model is retrained.
Zia can make two kinds of mistakes:
- An undesired image is approved by Zia: In this case, the user can remove the image.
- A desired image has been rejected by Zia: In this case, the reviewer can approve the image.
In both these cases, the users will have the option of letting Zia know that a mistake has been made. This will reduce the success rate.
If an image is allowed to stay in the system, Zia understands this to be a successful validation. If an image has been removed or approved but the option to notify Zia is not used, then Zia understands those images to be successful validations as well. The success rate will remain the same.
Training accuracy or Model score
The training accuracy or model score reflects the quality of your training data. The closer the training images are to the specified guidelines, the higher its training accuracy/model score. If you have multiple labels or objects, you'll be able to see the training accuracy score for each of them. Only rules with a training accuracy of over 80% will be used for validation.
Where can I find Zia Vision?
Users with the requisite permissions in their profiles will be able to access Zia Vision at
Setup () > Zia > Vision.
The following features are available in your Image Validation tab:
- New Rule: Use this button to create a new image validation rule.
- Image Validation Rules table: Use this to manage all your image validation rules. You will also be able to see useful details like the following:
- The module in which a rule is applied. You can filter the rules by module.
- The layout in which a rule is applied.
- The image field that is validated by the rule.
- The process used for validation (Match only, Detect only, Match and detect).
- The Model score which represents the quality of your training data. The rule will only be applied if its model score is above 80%.
- The status of the rule. You can filter by the statuses (All status, Active, Inactive).
Activate or deactivate a rule: Use this to toggle the status of each rule. Only active rules will be used to validate images.
To create an image validation rule
- Navigate to Setup > Zia > Vision > Get Started.
If you have created rules before, click New Rule.
In the Create Image Validation Rule page, enter the Rule name.
- Under the Where to validate section,
- Select the module from the drop-down list. This is the module where the image field is located.
- Select the required layout. For example, you could choose the Standard layout.
- Select the image field that needs to be validated.
You can either validate the record image, or a custom image upload field in a module.
Set the criteria. This could be either 'All records' or 'Selected records'. In case of 'Selected records', only images in records that satisfy that criteria will be validated by the rule.
Under the Validation Type section, select the type of validation you need for that image. This could be Match only, Detect only, or Match and detect.
- The options available in the Upload Training Data section will depend on the validation type chosen in the previous step. For each label or object, you can:
- Provide a name
- Choose whether the training images represent desired or undesired images in the case of labels. For each object, you need to choose if the images represent an object that needs to be detected or undetected.
- Add the training data from the Gallery or from your local device. If you are uploading from your local device:
- Ensure that the images have been moved or copied to a separate folder.
- Zip that folder.
- Upload the zipped folder.
- In case you are uploading images from your local device, you'll be able to:
- Add multiple labels or objects.
Upload multiple folders to training images for each label or object.
To learn more about this, refer to the section on uploading training data in this help document.
Under How would you like to add images to feedback learning?, you can decide what happens when an image rejected by Zia is approved by a reviewer. Select one of the following options:
- No, feedback learning is not needed: Choose this when you don't want Zia to be trained on images where it has made a mistake.
- Suggestions manually provided by Reviewers to users who manage rules, followed by the users' approval(s) to the suggestions accordingly: Reviewers will manually select the images to be added as suggestions for training. Users who manage rules will then pick from these suggestions and add them to the training set.
- Suggestions automatically provided by Reviewers to users who manage rules, followed by the users' approval(s) to the suggestions accordingly: Images that are approved by reviewers will be added as suggestions automatically. The admin will need to pick from these suggestions and add them to the training set.
- Click Save.
Note
- You can set an image validation rule for both standard and custom modules.
- Along with record images, only custom image fields with the Maximum images allowed property set to 1 are available for validation.
Whenever an image fails validation during record creation, the image will be sent for manual approval. The associated record will be created.
The image will wait under the My Jobs module to be approved or rejected by your reviewers.
Sometimes, you may feel that an image has been incorrectly approved by Zia. In those cases, you can remove the image from the record. When you do so, don't forget to select the checkbox in the popup that appears. This will help in producing an accurate success rate for the model.
- For a single image field, only one rule can be configured.
To test your image validation rule
Once the rule has been created, you can test the model that Zia will be using to validate incoming images. Based on this testing, you can tweak the model by modifying your training images. To test a model:
- Navigate to Setup > Zia > Vision.
- Select the rule for which you want to test the model.
Click Test the model.
In the Test the model popup, click browse or drag a file onto the popup to upload the image that needs to be validated by Zia.
Please note the supported formats and size limit mentioned in the popup before uploading your image.
Zia will validate the image and deliver the result.
- Click Test another image to repeat the process for another image.
To review invalid images
When an image is invalidated by Zia, it is moved to the My Jobs module for manual review. To review these images:
Navigate to My Jobs > Image Validation.
- For the record whose images you want to review, click the number of images waiting link.
- In the Image Validation Failure popup, click:
- Accept if image is valid.
- Remove if the image is invalid.
You can help Zia by pointing out if Zia made a mistake and/or by suggesting that image for training. Please note that this suggestion option will only be available if you've chosen to let reviewers manually suggest images for feedback learning.
- Click Save & Close.
To update an image validation rule
- Navigate to Setup > Zia > Vision.
- Hover over the rule that you want to edit.
- Hover over the (...) icon and select Edit.
- Make the necessary changes.
- Click Save.
Note
- You will not be able to modify the Module, Layout, Field, and Validation Type associated with that rule.
- You will be able to modify the Rule Name, Criteria for records, Training Data, and Feedback type.
- You will also be able to edit the rule by selecting a rule and clicking the Edit button.
To delete an image validation rule
- Navigate to Setup > Zia > Vision.
- Hover over the rule that you want to delete.
- Hover over the (...) icon and select Delete.
- In the popup that appears, click Yes, Delete.
Note
You will also be able to delete a rule by selecting a rule and clicking the Delete button.
Managing your training data
You'll need to add training data during the creation of an image validation rule, as well as when you feel that the performance of the rule could be improved.
To add your training data
Based on the validation type chosen for an image validation rule, you'll have different options under the Upload Training Data section.
If Match only is selected
- In the Upload Training Data section, select
- Desired if you want your records' images to match your training images. For example, if you have beautiful photographs of properties and you want your property records' images to match those.
- Undesired if you want your records' images to be unlike your training images.
For example, if you find that users sometimes upload images of the interior of a property as records' images. In this case, you can upload multiple images of the interiors of properties and select Undesired.
- Enter the label name for the set of training images that you're going to upload. For example, if you're uploading images of houses, then the label would be 'House'.
- Click Upload Image.
- In the Upload Training Data popup, you can either:
- Select the Gallery tab and select one of the available folders.
- Select the Desktop tab and upload one or more zipped folders containing custom images. For example, one zipped folder could contain images of the exterior of houses while the other could contain the same but for duplex units.
- Click Attach.
Note (only applicable in cases where you upload custom images)
- You can add additional custom images for a label by clicking +Images.
- You can add a maximum of three labels. To do this, click Add another label.
- Each validation rule will have one desired/undesired option. This will apply to all labels in that rule.
- Since labels are for entities that occupy at least 80% of the image, there can only be one label detected in an image. If multiple labels are present in a rule, the criteria pattern for the rule will always be Label 1 OR Label 2 OR Label 3. This cannot be modified.
If Detect only is selected
- Under the Upload Training Data section, enter the name of the object (Object name) in the training data.
- Click Upload Image.
- In the Upload Training Data popup, you can either:
- Select the Gallery tab and select one of the available folders.
- Select the Desktop tab and upload one or more zipped folders containing custom images of the object.
- Click Attach.
- Decide if you want the object to be detected or undetected.
- Click Add another object if you want to validate the presence or absence of other objects in an image. This option is only available when you upload custom images. You can add a maximum of three objects.
- If you have multiple objects, set the criteria for validation as well. For example, you may want the image to contain either a television or a combination of a table and a vase. In that case, click Edit Pattern, enter the criteria pattern, and click the tick icon.
Note You need the specific object to be crystal clear in every image you upload as part of the training data. For example, if you want to upload images of sofas, ensure that each training image only contains a sofa and nothing else. If the object will be present in multiple angles in different records' images, provide training images of the object in as many angles as possible.
The following are valid training images for detecting a sofa.
The image shown below is not recommended for detecting a sofa.
If you want to upload images from your desktop, you must follow these guidelines to achieve the best results from Zia:
Take a look at the cases below that highlight possible problems and the ways to troubleshoot them: