Written by Frida Kóbor, Tamás Matuszka / Posted at 10/2/25
Accurate, fast, low-cost, and automated: the aiData Auto Annotator delivers the quality you need, faster than ever
When it comes to ADAS development, high-quality annotated data is essential – but getting it shouldn’t be slow or expensive. Traditionally, manual annotation can cost millions of dollars annually and take months to process, delaying progress.
That’s why we created the aiData Auto Annotator at aiMotive: a powerful, automated solution designed to dramatically reduce annotation costs and accelerate processing times. With the recorded drives ready, our tool ensures that annotated datasets are delivered quickly and reliably – so researchers and engineers can keep improving the ADAS stack without delay.
Over the past few years, the Auto Annotator has helped us save tens of millions of dollars in annotation costs, and today, our ADAS stack is trained entirely on automatically annotated data – and the results speak to themselves. Our software stack outperforms many of the competing solutions, despite their substantially higher development budgets. Our annotator algorithms have achieved superhuman performance, surpassing even manual annotation benchmarks.
In the second installment of our new technical blog series, we’ll explore the transformative benefits of auto annotation from multiple angles, and for those curious about the inner workings of the tool, we included a dedicated section at the end that breaks down the technical details. Discover how you can maintain top-tier data quality without blowing your budget on manual annotation.
1. Automated 3D annotations for dynamic and static objects
The aiData Auto Annotator goes beyond conventional solutions by generating 3D bounding boxes for dynamic traffic participants – such as vehicles, pedestrians, cyclists, traffic signs and traffic lights, – and automatically labeling static elements like lanes, and road markings.
While other automated annotation tools exist, they still mostly provide 2D outputs, which limits the usability of the data for performing perception in 3D space. While 2D annotations are sufficient for training 2D object detectors, they fall short in accurately estimating the spatial properties of moving vehicles—such as distance and precise dimensions. In contrast, 3D annotations provide a richer, more accurate understanding of the driving environment, capturing the exact dimensions, orientation, and position of each object. Moreover, it provides depth information which is critical for precise localization and reliable decision-making. Additionally, 3D annotations enable occlusion handling, allowing systems to detect and interpret partially hidden objects – something 2D annotations struggle with. The result? A more robust and intelligent perception system that’s ready for real-world complexity.
2. Superhuman accuracy
Manual annotation has long been regarded as the gold standard in data labeling, with auto-annotation algorithms struggling to match its precision. However, manual labeling is not without limitations, most providers only guarantee around 95% accuracy. Achieving higher precision typically demands increased resources for additional rounds of quality checking and significantly higher costs, making it a less scalable solution for large datasets. But this era is over with the aiData Auto Annotator.
Our cutting-edge solution has not only caught up – it has surpassed manual annotation in precision and recall across all Operational Design Domains (ODDs). This breakthrough marks a pivotal shift in our data pipeline. We now almost exclusively rely on our auto annotation algorithms to generate training and validation datasets for our in-house AD stack. The result? Faster, more scalable, and consistently high-quality data – without compromising accuracy.
3. Instant effect on development budget
Annotation, if done by humans, is often the most expensive and time-consuming step in the development cycle of an ADAS/AD solution. Companies spend millions of dollars every year outsourcing this task, only to receive labeled recordings at a high cost and slow turnaround.
The aiData Auto Annotator dramatically reduces annotation costs, delivering a return on investment after just 30 hours of annotated recordings. Yes, the Auto Annotator require robust hardware. However, even factoring in additional sensors and GPU procurement, the savings can be 90%+ after 500 hours of annotated data - easily reaching eight figures in USD per year. A typical data collection campaign often requires tens of thousands of hours of recorded data, so the Auto Annotator could contribute to substantial savings in the development process.
Why spend your budget on human labeling when you could invest in what truly matters – enhancing your models, empowering your teams, and accelerating innovation.
4. Fast and scalable performance
In ADAS/AD development, time is everything. Manually annotating complex scenes could take weeks or even months, delaying critical decisions and slowing down innovation.
With the aiData Auto Annotator, those delays are a thing of the past. Faster access to labeled data means teams can begin training, testing, and iterating models almost immediately after data collection. Our solution delivers near real-time performance, thanks to a horizontally scalable pipeline that processes sequences in parallel across nodes of a cluster. That means large-scale auto-labeling happens fast. Really fast.
Why wait weeks for annotated data when you can start using it within hours of recording? Accelerate your development cycle, iterate faster, and stay ahead of the curve.
5. Highly adaptable for new use-cases
The aiData Auto Annotator isn’t just powerful; it’s highly versatile. We've successfully applied it across diverse locations and use cases, from drives in Japan to adaptations for trucking, consistently proving its reliability and flexibility in real-world scenarios.
What makes this possible? The Auto Annotator leverages powerful foundational models together with few-shot learning (FSL) and similarity search to refine existing categories and uncover new image classes with minimal additional data. In many cases, just a handful of images is enough to adapt the pipeline to a new domain.
This means the Auto Annotator can be adapted to your own unique use case quickly and effortlessly, without the need for extensive retraining or manual intervention.
Conclusion
The aiData Auto Annotator is redefining the standards of data labeling for ADAS/AD driving development. With automated 3D annotations, superhuman accuracy, massive cost savings, and near real-time performance, it empowers teams to move faster, smarter, and more efficiently than ever before. Ready to unlock your data pipeline?
This blog is part of a series covering how to build a scalable, reliable, and high-fidelity automotive data pipeline. Stay tuned, and if you’re struggling with any part of this, get in touch. We’re happy to help.
Under the hood – How the aiData Auto Annotator works?
The aiData Auto Annotator utilizes data from cameras, LiDAR, GNSS/INS, and optionally radar, and simultaneously annotates all sensors with a consistent 4D (space + time) environment model. However, to unlock its full potential, precise calibration and synchronization are essential (as discussed in our previous blog post) – similarly to manual annotation where these factors greatly influence the annotation quality. One critical factor that often gets overlooked is the accuracy of egomotion, the vehicle’s own movement through space. Since the Auto Annotator relies on an aggregated point cloud, even minor errors in egomotion can distort the size and position of bounding boxes, leading to flawed annotations.
The Auto Annotator is a hybrid solution that merges neural computing with traditional geometric and tracking methods to deliver precise and consistent annotations. First, a multimodal neural network followed by a non-causal tracker processes the sensor data to produce initial 3D bounding boxes. Then, the system generates an aggregated point cloud, built from sensor data and egomotion, which forms the spatial foundation for the annotation pipeline. Auto Annotator processes the sensor input and builds a temporally coherent world based on the aggregated point cloud by combining the outputs of the multimodal 3D object detection network, image and point cloud segmentation networks, as well as a set of computational geometry and tracking algorithms. Together, these components form the backbone of the Auto Annotator pipeline, resulting in annotations that are ready for tasks like model training or validation.
Our engineers also had to solve many challenges during the development of the pipeline to ensure that it provides the required accuracies. To give illustrative examples, compensating for the point cloud fragments recorded by a car speeding at 130 km/h, using a rotating LiDAR and a camera with a rolling shutter is definitely a challenge we had to solve. A rotating LiDAR captures data in slices as it spins, not all at once. At high speeds like 130 km/h, this causes distortions, especially when scanning fast-moving objects like oncoming cars, where relative speeds can reach 260 km/h. Similarly, cameras capture images line by line with a rolling shutter, not in a single instant. As the vehicle moves during capture, this also leads to image deformation. Compensating for these distortions is no easy task, but our engineers succeeded. We solved this issue by utilizing egomotion compensation and carefully taking the camera exposure time into consideration.
Another example is the challenge of filtering out ghost reflections, those bouncing off vehicle windows, storefronts or puddles. This is particularly tricky, because reflective surfaces can create realistic-looking ghost objects in the point cloud, mimicking nearby vehicles. Our teams have put a great amount of work into removing these ghost artifacts, ensuring that phantom detections caused by sensor limitations don’t make it into the final output.
The aiData Auto Annotator is a sophisticated system built over years to deliver highly precise results, but you don’t need to worry about the complexity behind it - we take care of making it all work seamlessly.