Answers are grounded in this paper and cite their sources.
Small gains in maritime object detection require repeated training and matched controls. We evaluated four detectors on SeaDronesSee (SDS), WSODD, and FloW, and examined training modifications using paired seeds, matched clean copies or views, and five synthetic degradation levels. SDS provides auxiliary validation-based evidence; WSODD and FloW provide held-out test evidence. Across six WSODD seeds, fixed single-view augmentation changed mean average precision (mAP50–95) by −0.10 percentage points (pp), with individual changes from −1.81 to +1.48 pp. Matched comparisons found no statistically supported gain attributable to degraded content under normal inputs, and the complete configuration showed no detectable difference from the scratch reference. Under the strongest synthetic low light, fixed single-view augmentation yielded +4.99 pp, positive in all six seeds. This benefit was not reproduced on the real low-luminance subset. Overall, no tested modification produced a statistically supported overall-accuracy improvement under normal inputs. Reliable assessment requires paired uncertainty estimates, matched controls, and explicit separation of synthetic robustness from evidence of real-world benefit.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: