Case Study: Leading Defense-Focused Research Company achieves 30% model compression and reliable stratospheric vessel detection on low-energy FPGAs with Multiverse Computing

Stratospheric Object Detection with Compressed YOLOv8n

Leading Defense-Focused Research Company engaged Multiverse Computing to detect vessels in aerial imagery captured from stratospheric balloons while meeting a strict constraint: run the model on low-energy, low-memory FPGA platforms. The primary challenge was fitting modern object-detection networks into constrained DPU memory without degrading performance.

Multiverse Computing implemented a compressed YOLOv8n pipeline using Singularity deep learning together with compression and quantization techniques, trained on the ShipRSImageNet dataset to target FPGA deployment. The result: up to 30% model compression (size reduced from 6.3 MB to 4.2 MB) while maintaining stable detection accuracy across image degradation (mAP@IoU > 0.67), enabling practical low-power FPGA operation.