Our client is an early stage is an autonomous micro mobility startup working on an autonomous delivery vehicle that can deliver small packages. They are based out of San Francisco, USA. They develop autonomous delivery robots that can deliver food and other essential supplies.
Their computer vision technology is designed to navigate primarily on sidewalks automatically detecting obstacles, recognizing road and traffic signs to deliver the package to the customer at the expected delivery time. The COVID-19 scenario has resulted in the demand for contactless delivery services our client is addressing this need.
Our client had hours of sidewalk video and thousands of individual frames to label. To achieve high levels of precision and accurate results, our client's data science team had to feed their neural networks with an immense amount of accurately labeled training data.
Data annotation 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 small team of engineers and designers.
They wanted to outsource the image annotation part. Naveen, the CEO of the company got in touch with us looking for a speedy and economical service.
We had to do full image semantic segmentation for tens of thousands of images as per the classification in the Cityscapes public dataset. The Cityscapes dataset focuses on semantic understanding of urban street scenes and after a quick analysis of the dataset, we decided that we had to annotate every image with about 15 classes.
We began with studying the dataset and preparing a training document for labeling the objects in the right classes and labels. We also worked with our client's team and made a study on the context in which the delivery vehicle operates and included information about instances that can be ignored in the training document.
We designed a specific process for this and divided our team into 5 small teams who were asked to label only 3 specific classes. Every time the 3 classes were labeled in a set of images, it went through a quality check before the other team started working on it. And after all the classes were labeled, there was a final quality check.
This way we checked the dataset for about 6 times in total ensuring that we had not missed any objects and all the objects were in their correct classes
Some metrics of the project
With our high quality annotation services, our client was able to improve their model and the client was extremely happy with the results.
Contact us for a free work sample.Contact Us