Peter Sunna, VP of product at Kognic, explains how the Embeddings feature of its data management platform enables customers to more quickly explore data sets, identify clusters and highlight outliers. The company will be at ADAS & Autonomous Vehicle Technology Expo 2024 in Stuttgart, Germany, to meet like-minded teams that are responsible for planning, developing and productizing ADAS/AD solutions.
Can you describe your company?
Kognic is the leader in data set management that enables global enterprises to assemble efficient ground-truth data pipelines for sensor fusion. The Kognic Platform has become the core software solution in the fields of advanced driving assistance systems and autonomous driving. It is currently being used by technology leaders such as Qualcomm, Bosch, Continental, Kodiak and Zenseact, which provide ADAS/AD systems that power global OEMs such as BMW, Ford and Volvo Cars. Supported by our industry-leading annotation engine, Kognic has critical tooling such as multisensor fusion, data exploration, pre-annotations and performance analytics that have been proved in many ADAS/AD deployments.
What will you be highlighting at the show?
Data exploration plays a crucial role in developing accurate machine learning algorithms. For all of us in the ADAS industry, we need to make it easier to navigate these complex data sets and gain a better understanding of which objects, scenes and sensors have an outsized impact on model performance.
“Through Embeddings, Kognic helps visualize these problematic clusters to accelerate data set optimization”
We’ll dive into the power of latent space visualization with embeddings. Embeddings represent complex objects in a computer-friendly format, enabling similarity searches and analysis of 2D representations. By using embeddings, engineers can more quickly explore their data sets, identify clusters and highlight outliers to improve their models.
We’ll also discuss pre-annotations at the event, and how they are best used to drive more efficient data pipeline workflows.
Can you give an example of how both products can help model training for ADAS/AV?
In a first implementation for a customer who is developing a deep learning ML model, enabling the exploration of latent space via embeddings allowed them to discover things about their sensor network and to find issues they didn’t previously know about. One example was optimization of how they handle objects behind windows or inside buildings. Through embeddings, Kognic helps visualize these problematic clusters to accelerate data set optimization.
In another customer application, focused on cuboids, using pre-annotation reduced average annotation time by up to 62% compared to fully annotating from scratch. The automation provided suggested positions for objects using dashed cuboids, enabling faster labeling. Although there was a slight reduction in the quality of annotations, with annotators putting less effort into identifying missed objects, the overall precision remained high. As just one example of the variety of shapes pre-annotations can be applied to, cuboid pre-annotations have great potential for accelerating the correction process for 3D cuboids, especially in highway scenes.
Who do you hope to meet in Stuttgart?
Kognic is excited to meet more teams who are responsible for planning, developing and productizing ADAS/AD solutions across all OEM and supplier segments. It takes a cohesive end-to-end effort to get it right, on time and on budget. We’ll share what we see that works and listen to what ongoing (and new) challenges are out there.
You can visit Kognic at booth 6618 at ADAS & Autonomous Vehicle Technology Expo, taking place at Messe Stuttgart, Germany, on June 4, 5 & 6, 2024.
This article first appeared in the April 2024 issue of ADAS & Autonomous Vehicle International.
