Multiverse Computing partnered with a European defense technology organization developing UAV and robotic solutions to bring advanced computer vision capabilities directly onboard tactical drones. By compressing and optimizing a state-of-the-art segmentation and tracking model, the project removed the dependency on ground compute and unlocked real-time autonomy on the drone itself, even under tight memory, energy and bandwidth constraints.
The Challenge
The client needed to deploy high-accuracy segmentation and object tracking on tactical drones operating with restricted compute, memory and power budgets. While the baseline EdgeTAM model delivered strong segmentation performance, its most demanding modules (the Spatial Perceiver, responsible for more than 75% of total inference time, and the Memory Attention module) created severe latency bottlenecks on embedded defense-grade SoCs. Real-time inference, with no sacrifice in segmentation fidelity, required a fundamental rework of those modules.
Our Solution
Multiverse Computing used its proprietary optimization technology to compress and optimize the EdgeTAM model for onboard deployment on the client's embedded SoCs. The work followed a three-step process: detailed profiling of the model on target hardware, tensor-network-based compression of the two bottleneck modules, and delivery of a production-ready model packaged for in-drone deployment.
- Profiling of RAM, latency and throughput on the target SoCs to identify the critical blocks.
- Tensor-network-based compression of the Spatial Perceiver and Memory Attention modules.
- Hardware-aware optimization tailored to the client's defense-grade embedded processors.
- Delivery of ONNX and TensorRT artifacts ready for in-drone deployment, with full benchmarking of latency, memory, energy and throughput.
Results
Real-time segmentation and tracking running fully onboard the drone
No reliance on ground compute or external link
Major latency reduction in the bottleneck modules
Lower energy consumption per inference cycle
Multiple AI functions in parallel within the same hardware budget
ONNX and TensorRT artifacts ready for in-drone deployment
Strategic outcome: a defense-grade UAV computer vision capability that runs end-to-end on the drone, preserves segmentation fidelity, and frees the operator to run additional AI functions inside the same hardware envelope.
