STELLAR Model Sets New Benchmark in Autonomous Driving Perception
A new 3D perception model named STELLAR has been unveiled by researchers, achieving top-tier performance in the Waymo Open Dataset challenge for autonomous driving. Utilizing a Sparse Window Transformer, this model combines inputs from LiDAR, radar, cameras, and map data. It has been trained on 50 million driving scenarios and possesses up to 500 million parameters. The research thoroughly examines scaling trends, linking the model's performance to its size, data, and computational resources. STELLAR surpasses previous approaches, illustrating that scaling strategies can effectively enhance perception in autonomous driving, despite difficulties related to sensor fusion and understanding 3D spatial contexts.
Key facts
- STELLAR is a 3D perception model for autonomous driving.
- It is based on Sparse Window Transformer architecture.
- Input modalities include LiDAR, radar, camera, and map prior.
- Trained on 50 million driving examples.
- Model has up to 500 million parameters.
- Achieves state-of-the-art on Waymo Open Dataset challenge.
- Outperforms prior arts in autonomous driving perception.
- Study reveals empirical scaling trends between model performance and size, data, and compute.
Entities
Institutions
- Waymo Open Dataset