GT Studio
  • Introduction
  • Getting started
    • Set up a Project
    • Set up Annotation Specification
    • Set up a Workflow
    • Create Jobs
    • Export Jobs
    • What's next?
  • Tools for Labelling
    • Image Segmentation
      • Tool Introduction
      • Maker Tool
      • Reviewer Tool
      • QC Tool
    • 2D Video and Image Labeling
      • Tool Introduction
      • Maker Tool
      • Reviewer Tool
      • QC Tool
  • Workflow
    • Workflow Steps
      • Steps and Patch
      • Step Analytics
      • Move and Push Jobs
      • Job Assignment
    • Job Build Structure
      • Base type - Image
      • Base type - Segmentation
      • Base type - Video
    • Workflow Routes
  • How to Guides
    • How to set up an image or a video classification task?
    • How to setup a Pose Tracking Project?
    • How to create jobs with pre-labeled data?
    • How to export annotation data in COCO format?
    • How to convert Playment segmentation mask to a grayscale mask
    • How to split a video into frames and create jobs in GT Studio
    • How to add classes after setting up the workflow
    • How to re-open a completed batch for making changes
  • Batches
    • Job Viewer
    • Quality Check
  • Annotator Performance
    • Video and Images
    • Segmentation
  • Annotation Specification
    • Classes
    • Attributes
  • Team management
    • Invite your team
    • Groups
  • API reference
  • Secure Attachment Access
  • What's New?
  • Hybrid Cloud
  • We are phasing out our SaaS offering — GT Studio
Powered by GitBook
On this page
  • Generate a QC
  • 1. Create Submissions
  • 2. Generate a QC Task
  • QC Results
  • 1. QC Metrics
  • 2. QC Details

Was this helpful?

  1. Batches

Quality Check

PreviousJob ViewerNextVideo and Images

Last updated 3 years ago

Was this helpful?

Once the annotators complete some of the jobs in GT create and these jobs reach the end step, you can perform a final quality check on a random sample from a batch.

‌To generate a QC sample, you first need to create a submission.

Generate a QC

1. Create Submissions

You can create a submission from the Final Stage tab in GT-manage. It shows a list of all the completed jobs grouped by date.

2. Generate a QC Task

a) Sample size: Percentage of jobs to sample from the batch

b) Frame Intervals: It is only applicable for Video projects. For longer sequences, we suggest keeping high frame intervals and a high sample percentage to ensure the sample represents many sequences.

c) Reviewers: Assign reviewers who will be performing the QC

After generating the sample, you can perform the QC task in GT create. To know how to perform a QC task, you can have a look here:

QC Results

As the sample tasks get solved, the result updates in the Latest QC section.

1. QC Metrics

Class Precision=Total images with zero critical class mislabelledTotal ImagesClass\ Precision = \frac{Total\ images\ with\ zero \ critical\ class\ mislabelled }{Total\ Images}Class Precision=Total ImagesTotal images with zero critical class mislabelled​
Geomtric Precision=Total images with zero critical geomety mistakesTotal ImagesGeomtric\ Precision = \frac{Total\ images\ with\ zero\ critical\ geomety\ mistakes}{Total\ Images} Geomtric Precision=Total ImagesTotal images with zero critical geomety mistakes​
Instance Precision=Total images with zero instance precision mistakesTotal ImagesInstance\ Precision = \frac{Total\ images\ with\ zero \ instance\ precision\ mistakes }{Total\ Images}Instance Precision=Total ImagesTotal images with zero instance precision mistakes​
Recall=Total images with zero critical missed objectTotal ImagesRecall = \frac{Total\ images\ with\ zero\ critical\ missed\ object}{Total\ Images} Recall=Total ImagesTotal images with zero critical missed object​
Precision=∑tTPt∑t(TPt+FPt)Precision = \frac{\sum_{t}TP_t}{\sum_{t}(TP_t+FP_t)}Precision=∑t​(TPt​+FPt​)∑t​TPt​​
Recall=∑tTPt∑t(TPt+FNt)Recall = \frac{\sum_{t}TP_t}{\sum_{t}(TP_t+FN_t)}Recall=∑t​(TPt​+FNt​)∑t​TPt​​

True Positive: If class, shape, tracker-ID, and all the attributes are correct

False Positve: If any of the above is incorrect for an annotation

False Negative: Annotation missed

2. QC Details

To improve the quality of the data, you can analyze the mistakes made by your annotators. Click on View QC details to check:

  • Class-Level accuracy

  • Mistake-Type distribution

  • Time spent in the QC by the Reviewers

Note: In video annotation, an object in one frame counts as one annotation. For example, a car present in ten frames counts as ten annotations.

🧠
QC for Image Segmentation
QC for 2D Video & Image Labeling