Multiverse Computing worked with a European defense technology organization to further optimize EdgeTAM for onboard computer vision on tactical drone platforms. Building on an earlier compression phase focused on the Mask Decoder and Memory Attention modules, the second phase expanded the optimization scope to include the Vision Encoder and Memory Encoder/Fuser, enabling a deeper reduction in model size while preserving most of the model's segmentation quality.
The Challenge
The client required a deeper compression of EdgeTAM to bring real-time segmentation and object tracking onto tactical UAV platforms with tight memory and compute budgets. A previous compression phase had targeted the Mask Decoder and Memory Attention modules. The challenge for this second phase was to expand the optimization scope to the heavier Vision Encoder and Memory Encoder/Fuser, while preserving most of the model's segmentation quality and improving inference performance on CPU.
Our Solution
Multiverse Computing applied a sensitivity-driven compression strategy to EdgeTAM, module by module rather than uniformly. Less sensitive modules such as Memory Attention and Memory Encoder/Fuser were compressed more aggressively, while more sensitive components such as the Vision Encoder and Mask Decoder were compressed more conservatively to control accuracy degradation. The optimized Multiverse Computing model is delivered with detailed module-level benchmarks and CPU performance gains for downstream onboard integration.
- 31.1% overall parameter reduction (from approximately 13.9M to 9.57M).
- Module-level reductions: 57.65% Memory Attention, 48.71% Memory Encoder/Fuser, 31.15% Mask Decoder, 15.85% Vision Encoder.
- J&F score of 60.6 versus a 64.3 baseline on the SA-V validation benchmark, approximately 94% relative retention.
- Approximately 19% CPU end-to-end speedup and approximately 25% higher frames per second.
- Compressed checkpoint ready for the next integration phase on target onboard hardware.
Results
Real-time segmentation and tracking running fully onboard the drone
No reliance on ground compute or external link
31.1% overall parameter reduction across EdgeTAM modules
Approximately 19% CPU end-to-end speedup
~94% relative J&F retention on the SA-V validation benchmark
Compressed EdgeTAM checkpoint ready for the next integration phase
Strategic outcome
A sensitivity-driven, module-level compression of EdgeTAM that delivers a 31% smaller checkpoint with approximately 94% segmentation quality retention and clear CPU performance gains, bringing high-quality segmentation and tracking closer to onboard UAV deployment.
