Storelift is a deep tech startup working on computer vision applications for retail. They are developing an autonomous cashier system that allows retailers to offer a checkout-free experience to their customers in store.
Storelift’s computer vision technology is designed to automatically track in-store clients and recognise products during the shopper journey to attribute the right product to the right customer.
To achieve high levels of precision and accurate results, the Storelift data science team has to feed their neural networks with an immense amount of training data.
Storelift had generated hours of shopping scenes in-store and split them into individual frames. They then laboriously annotated each person and each product in each frame. By doing so, they were able to map the coordinates of a person or product on an image. It would then indicate to the algorithm that an area on the image is of a person or a product.
This is an important but extremely time-consuming part of their data science team’s responsibilities. The process can take up to 70% of their time and often takes away the team’s attention from other important tasks. Annotating the training data in-house was cost-ineffective and tiresome for the team.
Together, we discussed which solution would best fit their needs. Storelift needed accurate annotations to get optimal performance and a quick turnaround time. They insisted in particular on the annotation precision. Even a few small errors can generate too much noise and dramatically reduce the algorithm prediction score.
At Sequence, our in-house technology allows us to work quickly and adapt to the needs of our clients. With a few adjustments, we were able to provide a project-specific annotation interface that fit their needs. This allowed our team of contributors to have more detail when annotating the images and gain more overall accuracy.
Within a couple of weeks, we were able to annotate over 50,000 images with roughly 100,000 annotations across all images whilst providing high-quality results and excellent accuracy.
Curious about our image annotation interface? See how it runs here: Annotation Tool Demo.