Six-Camera 3D Object Detection on Axelera Metis at 10 FPS

Six cameras, 3D object detection, and a chip about the size of my thumb. I wanted to see how far I could take that combination on Metis.

BEV perception is a demanding problem. Transformer-based systems and camera-to-BEV feature pooling have pushed the field forward, often with substantial compute and carefully optimized GPU kernels. I took on a different deployment challenge: get a strong, modern six-camera detector running on Axelera Metis, a tiny edge accelerator, and make the whole pipeline work as one system.

The motivation was practical. For ADAS systems with tight hardware budgets, the cost of the compute platform can determine whether an approach is feasible at all.

The hard part was deciding what should run where. I split the detector into four INT8 CNN stages on Metis, with the C++ host handling image preparation, camera geometry, feature lifting, resize operations, and final 3D box decoding. Each of the six camera views goes through the shared image network before their features are combined into a bird’s-eye view.

The result is a complete six-camera pipeline averaging about 10 frames per second, including image reads. The detections look strong across all six views, and the system is already running faster than the 2 Hz keyframe cadence of the nuScenes footage used in the demo. That gives me a solid base to build on.

Next, I’ll explore batching, scheduling across all four Metis cores, and pipelining across frames to push performance further. Seeing six-camera 3D detection running on a thumb-sized accelerator is a rewarding step towards surround-view perception for systems where space and hardware cost matter.

Code and demo: https://github.com/sanket-pixel/bevdet-metis