Improving mAP by 28 percentage points for an Asian OEM using 200,000+ multi-sensor annotated datasets from African urban analogs.

Client Overview

A leading Asian automotive manufacturer developing Level 4 autonomous vehicles needed high-quality training data for dense, chaotic urban environments. African cities like Lagos and Nairobi provided ideal analogs for their target markets, but lacked structured sensor datasets.

Project Scope

DataLens annotated 200,000+ frames from LiDAR, radar, and 8-camera arrays mounted on vehicles in Lagos and Nairobi. Tasks included:

• •3D bounding boxes for 50+ object classes
• •Semantic segmentation for drivable surfaces
• •Temporal tracking of erratic road users

Data was synchronized across sensors with sub-millisecond precision.

Challenges

• •Unpredictable behaviors: okadas, danfos, street vendors, and livestock.
• •Extreme lighting and weather variability (harmattan, rain).
• •Scalable annotation for 3D point clouds at 10 Hz.

DataLens Solution

Using the DataLens AI Studio, a team of 35 annotators (including traffic engineers) applied multi-stage validation. Custom tools enabled 3D annotation in real-world coordinates, with AI-assisted pre-labeling reducing manual effort by 60%. Field teams verified edge cases on-site.

Results

• •Improved mAP@0.7 from 52% to 80% on internal benchmarks.
• •Reduced false negatives for vulnerable road users by 41%.
• •Delivered in 12 weeks, accelerating the client's Level 4 pilot in Bangkok.