Multiverse Computing partnered with a consortium of aerospace and technology companies to bring real-time object detection capabilities to stratospheric platforms operating at the limit of power, memory and connectivity budgets. By compressing a state-of-the-art detection model to fit inside dedicated FPGA memory, the project turns high-altitude platforms into autonomous observation assets and removes the dependency on cloud or ground processing for the inference step.
The Challenge
Deploying real-time object detection from stratospheric platforms required extremely low power and memory consumption while maintaining reliable accuracy. The target hardware, FPGA-based and fully on-device, imposed tight constraints that standard detection models could not meet. Without compression, the model architecture was too large and too resource-intensive to fit into dedicated FPGA memory, limiting real-time performance and forcing the platform to rely on cloud or ground compute for any vision workload.
Our Solution
Multiverse Computing applied progressive compression and quantization to a state-of-the-art object detection model (YOLOv8-Nano), fitting the full architecture into dedicated FPGA memory. Two compression profiles were evaluated and benchmarked across multiple data levels, giving the consortium a clean way to pick the right operating point between size, power and accuracy for each mission profile.
- Soft compression profile delivering moderate size reduction with near-original accuracy.
- Strong compression profile maximizing memory and power savings with only marginal accuracy degradation.
- Full benchmarking of mean Average Precision at IoU across data levels and compression profiles.
- Real-time, on-device ship detection on FPGA hardware with no reliance on cloud processing.
Results
State-of-the-art detection compressed into dedicated FPGA memory
Real-time ship detection running fully on-device
No reliance on cloud or ground compute
Two compression profiles to balance accuracy, memory and power
Stable detection accuracy across varying levels of image degradation
Reusable approach across other on-device vision workloads
Strategic outcome: a stratospheric observation capability powered by a high-accuracy object detection model that runs end-to-end on FPGA hardware, with two operating points to balance accuracy, memory and power for each mission profile.
