Problem
People in aerial disaster scenes are often only a few pixels wide, occluded by debris, and surrounded by textures that defeat standard detectors.
Contribution
- Identified the stride-4 P2 detection scale as the useful change, raising very-tiny recall from .743 to .757.
- Rejected a Mamba neck after it added 2.4M parameters and about 2.8× latency without improving accuracy.
- Raised very-tiny recall to .850 with 1280px test-time augmentation when a higher-latency operating mode is acceptable.
Case note 01
Ablate, don't decorate
Starting from YOLO11m, the study added CBAM, a P2 head, and a bidirectional state-space neck one at a time under one protocol. The contribution is as much the negative result as the winning model.
Case note 02
The useful change
The P2 scale operates at stride four, preserving spatial detail for sub-eight-pixel targets. It improved AP50 and very-tiny recall for about half a million parameters and one millisecond of latency.
Case note 03
Limits kept visible
Results use one semi-synthetic dataset and one seed per configuration. Desktop-GPU latency is not an airborne-device measurement; validation on real rescue imagery remains future work.