Back to projects
Jul 01, 2024
1 min read

Autonomous Driving Scene Analytics

Applied vision model fine-tuning and taxonomy-based retrieval to surface rare and critical driving scenarios.

Challenge. Rare driving events (e.g., pedestrians crossing unexpectedly, animals on the road) are hard to discover and analyze at scale.

Approach. Fine-tuned a multimodal vision-language model to detect anomalies. Generated structured scene descriptions and indexed them in a search pipeline using a domain taxonomy.

Outcome. Reduced discovery time for rare events from hours to minutes, improving the ability to evaluate and retrain ADAS systems on critical edge cases.