Budget, Operating Point and Prediction Fusion in Maritime Small-Object Detection: A Controlled SeaDronesSee v2 Study Using EVREN

Detecting small objects in high-resolution maritime UAV imagery requires balancing spatial detail against processing cost. Selective tiling directs additional detector calls to promising regions, but its benefit is difficult to interpret when competing methods use different resolutions, thresholds or fusion rules. We investigate these factors on SeaDronesSee v2 using six YOLO11n exports trained through EVREN. The study combines approximately matched-time comparisons, a source-video-separated sensitivity allocation, operating-point analysis and a controlled adaptation of MRU-YOLO. At approximately matched mean processing time, selective high-resolution tiling reduced small-object recall relative to exhaustive lower-resolution tiling by 1.739 percentage points on average across the three official-split exports. The subsequent source-video allocation showed the same direction. Learned crop selection slightly improved operating recall while reducing AP, and hybrid fusion reduced false positives at the cost of additional missed detections. On the official 3,750-image hidden test, the fixed selective policy achieved 47.17% AP compared with 48.12% for exhaustive lower-resolution processing, at local mean times of 194.08 and 198.37 ms/image, respectively. The results demonstrate the importance of evaluating computation, operating point, selection and fusion jointly when assessing selective inference. The contribution is a controlled empirical evaluation, supported by retained predictions and public replay materials. Findings remain conditional on the evaluated exports and allocations; prior development exposure and incomplete cloud-training provenance limit generalization. This is an experimental research preprint, not yet peer reviewed. Version 1.0.0 reports evidence available on 25 October 2026 at 21:54 TRT (UTC+3), including six of fourteen planned official-test results. The remaining eight results, including D-FINE accuracy, are reserved for a subsequent version. The deposit includes the full English article, an official-test results table and a README. Earlier public code and replay materials: https://github.com/mehmetaksoy/seadronessee-compute-audit/releases/tag/research-snapshot-2026-09-27 . That September snapshot covers the earlier manuscript, 81 completed conditions and three S01 official scores; it does not contain the new D-FINE implementation. The author acknowledges EVREN and the Secretariat of Defence Industries (SSB), Presidency of the Republic of Türkiye, for access to dataset-management and GPU-training infrastructure, and the SeaDronesSee/MaCVi contributors for the benchmark. The article states the scope of this support, AI assistance and author responsibility. CC BY 4.0 applies to the study-authored article, figures and aggregate results; dataset photographs and model weights are not redistributed. Upstream components retain their original licenses.