Reducing end-to-end latency for ultra-large aerial images

Lower the end-to-end latency of processing ultra-large aerial images to enable practical wide-area spatio-temporal scene understanding and real-time UAV deployment.

Background

The HARD benchmark uses 122-megapixel UAV frames, whose processing requires loading, tiling, per-tile detection, fusion, and tracking. The paper shows that this pipeline can require seconds per frame, making latency consequential for streaming evaluation through the proposed s-HOTA metric. The authors identify reducing this end-to-end latency as one of two open problems requiring deeper study.