MANTIS Neural Compression

MANTIS is a task-informed split-computing system for UAV perception. It compresses visual evidence before transmission by estimating task relevance on the client, modulating the neural analysis transform, entropy-coding the resulting latent, and routing the compact representation to task-specific edge heads.

The system is evaluated across UAV-relevant workloads:

• UAVid semantic segmentation for urban scene parsing
• WAID wildlife detection for aerial environmental monitoring
• Boreal Fire smoke detection for wildfire intelligence

Compared with JPEG, WebP, LADON, learned image codecs, and non-conditioned ablations, MANTIS improves the low- and mid-rate task-utility frontier. The paper reports up to 62.2% bitrate reduction at matched downstream accuracy and up to 9.3% average normalized task-accuracy improvement at matched bitrate.