Who is the annotator - your PC or yourself?

To make sure your model actually has something meaningful to learn from, there is one crucial step in any AI project: 𝗔𝗻𝗻𝗼𝘁𝗮𝘁𝗶𝗼𝗻.

Annotation is the process of telling the model what to look for and what it should recognize. In practice, it often means drawing bounding boxes around the objects you want the model to 𝗱𝗲𝘁𝗲𝗰𝘁 𝗮𝗻𝗱 𝗹𝗮𝗯𝗲𝗹𝗶𝗻𝗴 𝘁𝗵𝗲𝗺 𝗰𝗼𝗿𝗿𝗲𝗰𝘁𝗹𝘆.

The levels of annotation can be divided into different levels:
• Classification (motorcycle somewhere in the picture)
• Detection (box around motorcycle)
• Segmentation (motorcycle marked pixel by pixel)

It is also one of 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁 𝘁𝗶𝗺𝗲-𝗰𝗼𝗻𝘀𝘂𝗺𝗶𝗻𝗴 𝗮𝗻𝗱 𝗰𝗼𝘀𝘁𝗹𝘆 𝗽𝗮𝗿𝘁𝘀 𝗼𝗳 𝗺𝗮𝗻𝘆 𝗔𝗜 𝗽𝗿𝗼𝗷𝗲𝗰𝘁𝘀.

One way to reduce the cost is to reuse already annotated datasets or rely on open source tools where possible.

At Move, we have developed an internal Move Data Framework to support this process.
The framework is built to 𝗽𝗿𝗶𝗼𝗿𝗶𝘁𝗶𝘇𝗲 𝘀𝗽𝗲𝗲𝗱 𝗮𝗻𝗱 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗼𝗻, allowing as much work as possible to be done automatically before a human annotator performs the final verification.

Tools can help a lot with annotation, but in the end t𝗵𝗲 𝗾𝘂𝗮𝗹𝗶𝘁𝘆 𝘀𝘁𝗶𝗹𝗹 𝗱𝗲𝗽𝗲𝗻𝗱𝘀 𝗼𝗻 𝘁𝗵𝗲 𝘁𝗿𝘂𝘀𝘁 𝗮𝗻𝗱 𝗰𝗼𝗺𝗽𝗲𝘁𝗲𝗻𝗰𝗲 𝗼𝗳 𝘁𝗵𝗲 𝗮𝗻𝗻𝗼𝘁𝗮𝘁𝗼𝗿.