A fast-growing startup in the autonomous delivery sector was developing advanced object detection systems capable of reliably detecting objects on the road, such as traffic signs, pedestrians, and vehicles. Given the complexity of real-world environments, this required integrating multiple sensor modalities to ensure accurate and consistent object detection.

One of the biggest challenges was annotating the vast amounts of sensor data, including Lidar point clouds and camera images, to create a dataset that could capture the nuances of different driving scenarios. Traditional labeling methods made quality control across different sensors extremely time-consuming. The team needed a solution that could seamlessly integrate data from multiple sensors, enabling consistent object tracking across both sensors and time.

To tackle this, they ingested synchronized Lidar and camera data into the Mindkosh platform combined with pose information about the ego-vehicle. This combined data helped the labeling team easily distinguish between static and dynamic objects. The ability to view camera images alongside point clouds provided valuable context during labeling, allowing accurate labeling even in challenging conditions, such as poor lighting or occlusion. Further enhancement came from using RGB values from the camera images to color the point clouds, making it easier to distinguish objects.

Maintaining high annotation quality was critical, so the team implemented a rigorous quality control process using the platform’s built-in tools. A custom labeling workflow was set up, incorporating Honeypot and spot-checking methods to verify data quality. This approach significantly improved training data accuracy while also saving time.

To further reduce errors and maintain label consistency across point clouds and images, our team introduced an automatic error-checking mechanism. Labels from point clouds—both cuboids and segmentation—were projected onto camera images and validated using an estimation model based on Conditional Random Fields (CRF). This helped flag potential mistakes, which were then reviewed by human quality checkers.

By combining accurate annotations with sensor fusion, the startup achieved a noticeable improvement in model performance, particularly in challenging environments. The time spent on perforing Quality checks was reduced by over 40%, and the streamlined quality control process further eased the workload for the ML team. As a result, development timelines were accelerated without compromising data quality.

To know more about our Automatic error-checking model and how it can help you massively reduce your Quality checking efforts, get in touch with us, or just email us at hello@mindkosh.com