Post

Data on the Edge with Deepen: Key Takeaways

January 06, 2022

In the third episode of Data on the Edge, Jai Natarajan, VP Strategic Marketing and Business Development of iMerit, sits down with Mohammad Musa, Founder and CEO of Deepen AI, a leading LIDAR tool that iMerit partners with, to explore where in the ML DataOps cycle does Deepen AI provide value, edge case examples which the platform could overcome, and the state of LIDAR in the near future.

Here are 5 key takeaways from the session:

  • Goal of Deepen AI
    Deepen AI’s initial focus was to build safety tools for multi sensor systems and as one of the most important aspects of safety is validating real data, they built a full annotation suite for sensor fusion and then expanded to calibration of multi sensor and a safety validation suite for autonomous vehicles.
  • How Deepen AI helps in DataOps
    • Pre Data Collection – Calibration suite to ensure you are collecting the right data
    • During Data Collection – Algorithms that can be used in real time to detect major issues in calibration
    • Post Data collection – Annotation tools to visualize the data, annotate at very high accuracy, track that across different sensors and validate the data
  • Reasons for edge cases
    Edge cases can take place due to bad sensor data, around unseen situations, and things that you have seen before but are configured in an unpredictable way, for example traffic controllers asking to move in a certain direction.
  • Examples of edge case areas 
    • Detection: Is there stuff in front of me?
    • Classification: What am I looking at?
    • Segmentation: Where does the human end and car start? /Where does the road end and the sidewalk start?
    • Prediction: Understand and predict the movement of visible or occluded objects
  • State of LIDAR today and in the near future
    The most exciting thing about LIDAR is that it is now in production and the threshold of quality is being met. Different technologies within each LIDAR segment are also getting specialized. Detection, segmentation, tracking and prediction quality has gotten to a point where it is good enough for a lot of cases so the main focus soon will be on higher order understanding of scenarios which have a combination of things that one needs to be aware of.