DenseScout: Algorithm-System Co-design for Budgeted Tiny Object Selection on Edge Platforms

Deploying tiny object perception on edge devices like drones or cameras is notoriously hard due to limited compute and real-time demands. Traditional approaches use heavy offline detectors to pre-select candidate patches, but these often fail under strict budgets—strong offline accuracy doesn't translate to effective low-budget prioritization or usable performance once transport and inference delays are factored in. To address this, researchers at Zhejiang University present DenseScout, a lightweight dense-response selector with just 1.01 million parameters. It directly ranks candidate patch locations from high-resolution images using a lightweight proxy input, making it far more effective for low-budget tiny-object selection than detector-style frontends.

DenseScout also introduces a transport-aware runtime for heterogeneous edge devices like Rockchip RK3588 and NVIDIA Jetson Orin NX, along with a novel metric called QoS-constrained recall. This metric counts a target as successfully perceived only if it's covered by selected regions and the entire end-to-end processing finishes before a deadline. Experiments show DenseScout consistently outperforms detector-based baselines in offline budgeted patch-selection, especially at low budgets. The results underscore that edge tiny object perception must be optimized as an algorithm-system co-design problem, not just isolated model selection.