Interpolation is a process to estimate the location of labels within frames in videos and image sequences, using existing manual labels as reference points. Since only estimates are provided, all labels created via interpolation are assigned a confidence score (α) of 99%.
Interpolation can be used on several ontology shapes.
Since interpolation only uses label information, no data is stored on our servers.
Video Tutorial - Interpolation
|Rotatable bounding box||✅||✅|
- Manual labels need to be added to some frames within the range you would like to interpolate.
The more instance labels are manually added throughout the video, the more accurate the resulting interpolation will be.
- Click the Automated labelling button in the bottom left corner of the label editor to bring up the options for automated labeling
The 'Tracking and interpolation' section will be open by default - select the object instance(s) you would like to interpolate.
Under the 'Method' heading, select Interpolation.
'Interpolation range' specifies the range of frames that will be interpolated between.
Please note that only organizations on 'Enterprise' pricing can interpolate over a range larger than 1000 frames.
- Once you are satisfied with the settings click Run interpolation to begin interpolating the object instances.
Interpolation results can be improved by manually correcting some interpolated labels, and re-running the interpolation.
Enable the 'Interpolation auto adjustments' toggle in the 'Drawing settings' section of the label editor settings (seen in the screenshots below). This ensures only labels with a confidence score (α) of 99% will be overwritten in successive interpolations, while manual labels (α = 100%) are kept and serve as key frames.
This process can be repeated any number of times until all labels have the desired accuracy.
All labels created using interpolation will always be assigned α = 99% irrelevant of how often re-interpolation is run.
Updated about 2 months ago