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

On September 7, 2026, arXiv released DenseScout, a budget-oriented tiny target selection algorithm for edge platforms. This algorithm processes only the areas selected by downstream detectors through lightweight front-end proxy sorting, featuring 1.01 million parameters and eliminating box regression steps to directly optimize patch center priorities. DenseScout implements task-specific selector formulas, aligns output representations with supervision and decoding, and is jointly designed with perception transmission execution and quality of service evaluation. On VisDrone and DOTA datasets, it provides higher low-budget recall than detector-based selectors, and controlled fixed K inspection experiments prove its superiority over selective proxy baselines. Cross-platform tests show that its deployment effectiveness on Jetson Orin NX and RK3588 depends on the combined effects of selector quality, memory allocation, and heterogeneous runtime implementations.