Suggested Workflows

V7's workflows enable you to create a limitless number of custom training data pipelines. Here are some of our recommended workflows broken down by use-case to give you some inspiration.

External Annotation

If you're working with an external workforce like V7's Labelling service, set up a secondary review stage to review all or a percentage of the files that their team has annotated and reviewed

Example 1 - Standard external annotation flow

1. External workers add annotations
2. External reviewers review annotations against instructions
3. Internal reviewers perform a final review

Example 2 - Sampled external annotation flow

1. External workers add annotations
2. External reviewers review annotations against instructions
3. A 50% sample of files is sent for internal review
4. Internal reviewers perform a final review

Example 3 - Model-assisted sampled external annotation flow

1. Initial annotations are made using a model
2. External workers review model-created annotations, add missing instances, and correct model output where needed
3. External reviewers review annotations against instructions
4. A 50% sample of files is sent for internal review
5. Internal reviewers perform a final review

Specialist Reviewers

Example 1 - Specialist review routed by tags/classes

1. Workers add annotations including "Difficult" tag for cases requiring specialist review
2. Files are routed to the specialist review if they contain the "Difficult" tag or the generalist review if they do not
3. Reviewers perform a final review

Example 2 - Consensus review

1. Workers annotate blindly in parallel
2. Files are routed to the review stage if the annotations they contain do not meet IoU thresholds set by workflow creator. The reviewer manually discards incorrect annotations.
3. Rejected files are sent to a separate annotation stage.

External Automation

Example 1 - Custom automation flow

1. Workers add annotations
2. Annotation data/file metadata payload is sent via a webhook, triggering a custom automation (this can be a no-code solution like Zapier, or an Object Lambda Function). This automation can incorporate V7's SDK to add a comment/annotation/change an annotation stage
3. Failed files are sent to a separate Annotate stage where they can be filtered for and reviewed
4. Reviewers perform a final review

Document Processing

Example 1 - Receipt scanner flow

1. Text is pulled into bounding box/polygon fields using one of the Text Scanner/Invoice Scanner/Receipt Scanner models
2. Workers review model-created annotations, add missing instances, and correct model output where needed
3. Reviewers perform a final review

Dataset Optimisation

Example 1 - Remove corrupt data

1. "Blurry/Corrupt" tag is added to applicable files using a simple classification model
2. Files are routed to the annotation workflow or to an Archive stage if they contain a "Blurry/Corrupt" tag
3. Workers add annotations
4. Reviewers perform a final review

Example 2 - Identify abnormal data

1. Files are run through a custom script prior to upload which applies an "Outlier" tag to files that are dissimilar to the rest of the dataset.
2. Files are routed to the annotation workflow or to a Webhook stage if they contain a Blurry/Corrupt tag
3. "Outlier" files metadata is sent to an external database (can be as simple as a Google Sheet using a Zapier automation) so similar files can be sourced and added to the dataset
4. Workers add annotations
5. Reviewers perform a final review