Driving AI Readiness with Scalable Multiclass Vehicle Detection

About the customer

A leading geospatial analytics company leverages AI and satellite data to deliver actionable insights for global enterprises and governments. Their advanced platform transforms massive datasets-including satellite imagery, mobile signals, and IoT feeds-into clear, predictive intelligence that drives operational efficiency and strategic advantage. From infrastructure monitoring and supply chain visibility to population movement analysis, they enable faster, data-driven decision-making across industries, enhancing agility and real-world responsiveness.

Goals

Deliver high-quality, consistent ground truth annotations using 30 cm high-resolution electro-optical satellite imagery, enabling the development of a high-performance multiclass vehicle object detection model. The objective was to accelerate AI model development and deployment through fast turnaround, rigorous quality control, and cost-efficient labeling, reducing time to production and improving model accuracy at scale.

Challenge

The client needed high-quality ground truth labels for over a million data points, with a fast turnaround time and strict quality requirements. Existing solutions were either too costly, slow to scale, or lacked the precision required for mission-critical geospatial models. To develop a multiclass high-performance object detection model, precise annotations and attention to the details of the multiclass vehicle types in the remote sensing images are essential. Given the diverse environments, types of vehicles, and varying weather and terrain conditions present in these images, achieving high-quality and precise labeling is a complex task. The primary challenges include occlusions, vehicle orientations, various vehicle attributes, and distinguishing cars, vans, and trucks from similar-looking objects. The target objects should be annotated using a polygon tool, and necessary metadata should be added. It is also essential to ensure that the annotated labels do not overlap with adjacent objects.

Solution

Labelbees provided end-to-end data annotation through its human-in-the-loop pipeline, combining Labelbees’ ontology knowledge base, expert annotators, custom workflows, and robust quality control. The team delivered hundreds of thousands of labeled assets within weeks, enabling the client to accelerate model training and reduce development time.

The project required a detailed ontology study and knowledge of the target objects. Labelbees' in-house data labeling team has overcome the challenges with the help of in-house subject matter expertise, machine learning data specialists, and data scientists. Labelbees has captured over 1.3 million data points, including annotated polygons and metadata attributes. Labelbees has followed the high-standard data labeling workflow with custom services that include:

• Ontology Development & Review
• Edge Case Identification & Curation
• Custom Metadata Collection
• Detailed Labeling Guidelines & Documentation
• Specialized Labeling Strategies per Class
• Robust Quality Assurance Process
• Iterative Feedback Loop
• Dataset Insights & Reporting

Result

Labelbees' expertise in high-quality annotation contributed to the success of a state-of-the-art multiclass vehicle object detection model. The models achieved superior detection performance by ensuring precise and consistent labeling, making them more effective for several use cases, such as traffic management, urban planning, environmental monitoring, air pollution assessment, and economic analysis applications.

In our work, we enabled our customer with faster deployment of AI features by significantly improving internal efficiency across their machine learning pipelines.

• Achieved near-instant labeling delivery, cutting annotation cycles from 14 days to under a day
• Consistent 10-15% boost in annotation efficiency
• Avoided costly rework and delays
• Achieved first-pass approval
• Faster deployment of AI features to customers
• Reduced model iteration cycles
• Competitive edge in time-sensitive markets
• Improved internal efficiency across the ML pipelines