Data Curation

Pain Point

You know what you need to annotate, but you have a hard time extracting it from your pool of unlabelled data. You have to manually go through the data to select desired images. You wrote a dataset preprocessing script that’s hard to maintain because it’s not in your model development pipeline.

Solution

Curate provides you with the right filtering tools to find images based on your metadata search criteria. The process can be automated or semi-automated, as needed.

Pain Point

Your model performs poorly for some object classes so you’re forced to manually find instances of these classes. Worst case scenario: You trigger another data collection initiative to find more desired images, restarting your project and wasting more time.

Solution

Curate offers systematic ways to find images in your datapool that improve your model on desired object classes.

Pain Point

You have a lot of video footage, but there aren’t many frames which contain interesting objects.  You’re stuck sending the entire set of videos for professional annotation, or you assign someone to manually select video sequences for annotation based on intuition.

Solution

Curate can find the video sequences with the richest information so that annotators can focus on the most relevant frames for model training.

Pain Point

Your model performs well during its training phase but it has poor performance in production. You find no good tools on the market that speeds up the investigation of your training data vs your production data.

Solution

Curate helps identify key differences between two (or more) datasets. This is important because if your training data isn’t representative of your production data, your model won’t perform well.