Building Production-Ready ADAS: Zenseact's Journey to Efficient, High-Quality Annotations (Clone) (Clone) (Clone)

Industry

Autonomous driving software

Challenge

Zenseact ran into scaling issues with their annotation pipeline. Manual annotation at scale was slow and expensive. Keeping quality consistent across different teams and suppliers was tough. And their engineers were spending way too much time on manual quality checks instead of building product.

Results

Working with Kognic, Zenseact cut annotation time per sequence dramatically and boosted throughput. Engineering review time dropped significantly. They also gained transparent cost control tied directly to model performance, saving up to 46% through pre-annotations.

Key Product

Pre-label workflow with discount model, Auto-QC automation for quality assurance

Visit Zenseact
image (24)

About Zenseact

Zenseact is Volvo Cars' safety software partner, building the ADAS and autonomous driving features that power vehicles like the Volvo EX90 and Polestar 3. They're based in Gothenburg, Sweden, and they're known for pushing the boundaries of multimodal perception, camera, lidar, radar, backed by large-scale AI training on NVIDIA infrastructure. They also publish cutting-edge research and maintain the Zenseact Open Dataset (ZOD). For a company shipping safety-critical driver assistance systems, data quality isn't negotiable.

80%

Fewer objects processed due to the removal of already highly accurate classes.

10.5

MSEK saved through pre-annotations as of November 2025

1517733385655
Mathias Broxvall
Unit Leader DL Data Preparation @ Zenseact
Partnering with Kognic helped us scale high-quality annotations in DL 2.0 while evolving our workflows around pre-labels, review automation, and a better, more efficient way of working. The collaboration has been flexible and fast, which is critical for safety-grade perception.

The Challenge

As Zenseact ramped up ADAS development, their data preparation pipeline started to buckle. Manual annotation at scale was proving to be a major drag: slow, expensive, and hard to iterate on quickly. Quality consistency became a real issue when you're coordinating across multiple teams and external suppliers. One bad batch could ripple through to production and affect safety-critical perception systems.

On top of that, manual QC was eating up engineering time. Their team was spending hours reviewing annotations instead of developing models. They needed a way to keep quality high while dramatically speeding up throughput and cutting costs per annotated unit.

The Solution

Kognic and Zenseact worked together to put two key workflows into production.

Pre-Label Workflow: Instead of annotating from scratch, the teams built a system where model-generated annotations are imported into Kognic and verified by annotators. Annotators just validate or adjust what the model produced, which saves time even when edits are needed. The pricing model is tied to how much work the annotator has to do—full creation, adjustment, or simple verification—so costs scale with model performance.

Automatic Quality Control (Auto-QC) Implementation: Kognic built Zenseact's Auto-QC rules directly into the workflow to catch errors early. These rules are customized for each task to flag the most expensive mistakes before they make it downstream, preventing costly rework later.

The Results

The partnership delivered real, measurable results. Zenseact cut time per sequence significantly and increased throughput without sacrificing quality. The "discount ladder", where pricing is tied to how much annotator effort is required, gave them better cost control and led to savings of up to about 47% on dynamic objects and 38% on static ones.

Auto-QC made a huge difference on the quality side. Engineering review time dropped dramatically, freeing up the team to focus on core development. That meant faster acceptance cycles and fewer late-stage iterations, speeding up the whole development process while maintaining the safety standards production vehicles demand.