Challenge
An AV technology company expanding into APAC had a critical problem: perception models trained on Western datasets (Waymo, nuScenes) failed catastrophically on Indian road scenarios. Dataset gaps included missing vehicle classes (auto-rickshaws, cycle-rickshaws, hand-pulled carts, overloaded two-wheelers), different road structure (unmarked roads, construction zones, severe potholes), unusual traffic behaviour (mixed flows, animals on roads, pedestrians crossing mid-highway), and weather (monsoon rain, dust haze, night driving). The model mAP on Indian road scenarios was 54% - below any acceptable threshold.
Solution
Phase 1 Data Collection: Dashcam and LiDAR-synced video collection across 15 Indian cities. Diverse conditions: urban/rural, day/night, monsoon/clear, highway/city. 200,000+ frames captured. Phase 2 Multi-Modal Annotation: Bounding box annotation across a 45-class vehicle taxonomy including all India-specific classes. Semantic segmentation for road surface, lane markings, and drivable area. 3D cuboid annotation for LiDAR point cloud data. Keypoint annotation for pedestrian pose estimation. Full metadata tagging for scene conditions. Phase 3 Quality Assurance: Removal of blurred, overexposed, and corrupt frames. Label consistency cross-checks. Gold-standard validation with 20% honeypot images. Final annotation accuracy: 98.2%.
Results
Key outcomes: Model accuracy improved from 54% to 91% mAP - a 37-point improvement. The annotated dataset became the client primary training asset for APAC expansion. The India-specific 45-class vehicle taxonomy filled a gap that no public dataset had addressed. The annotation schema became a new internal standard for the client globally. The client accelerated their India market launch timeline.