A Conversation on Pre-Annotations with Tommy Johansson.
Tommy has been working in the automotive industry since 2006, specializing in ADAS/AD. With a background in both OEMs and Tier 1-2 suppliers, Tommy brings a wealth of experience in writing requirements, testing, and verification of products within the industry.
We recently had the opportunity to sit down with him to discuss the exciting world of pre-annotations. Pre-annotations are a powerful tool that can significantly improve the efficiency of annotation tasks in the field of computer vision. Here are some key insights from our conversation with Tommy:
Some believe an offline ML model with no real-time constraints can produce high quality ground-truth data. In our experience there are many situations that still need human feedback.
Tommy highlighted that unless you can trust your pre-annotations 100%, the main value relies in guiding where to focus. Pre-annotations can help indicate the presence of an object, although they may not always provide accurate information about the object's size or class. The goal is to strike a balance in terms of annotation thresholds, ensuring that annotators can efficiently adjust and refine pre-annotations without spending excessive time on unnecessary modifications.
One observation Tommy shared was that, in many experiments, annotations made from scratch using our efficient tool at Kognic still yield better results compared to pre-annotations. While pre-annotations can provide guidance to annotators as noted above, the process of micro-adjusting and making unnecessary adjustments to pre-annotations can often take more time than simply drawing annotations from scratch.
Tommy emphasized the ongoing need for human feedback in the annotation process. While the quality of pre-annotations may improve over time, human expertise will remain valuable in correcting annotations and making decisions in complex situations. As the technology advances, human feedback will likely shift towards higher-level expertise, providing guidance on critical decisions made by autonomous systems.
The obvious value of pre-annotations lies in the time-saving aspect. By providing pre-annotations, we can help annotators work more efficiently and deliver faster ground truth to our clients.
For our clients, getting started with pre-annotations involves running their existing models on the data they provide. If they do not have a model, our team can offer support and guidance. We then import the data with the help of our APIsn. For custom integrations to your cloud or data center, our Solutions Engineering expertise can help you navigate infrastructure and 3rd party services to get the most from your dataset. While the quality of pre-annotations may vary based on the specific use case, our efficient tools and expertise ensure that the annotation process remains effective and efficient.
As the field of pre-annotations continues to evolve, Tommy remains excited about the possibilities. By combining the strengths of our annotation tools and pre-annotations, we can deliver even better results for our clients. Continuous experimentation and methodological testing will drive further improvements in both efficiency and accuracy.
Stay tuned for more updates on pre-annotations and the groundbreaking work being done by Tommy and our team at Kognic!