If you write a maritime-detection paper today, your baseline is YOLO — a single-pass convolutional detector tuned for speed, which ships because autonomous navigation cannot wait on a slow pipeline. Hundreds of maritime YOLO adaptations exist (v4 through v11, with attention heads, lightweight backbones, feature-pyramid tweaks). The honest question the research rarely answers is not 'does YOLO detect ships?' — it does — but 'does it detect them when it matters, from where a ship's camera actually sits?'
This article reads the YOLO literature carefully: what the scores mean, which datasets they were measured on, and what is systematically missing from the papers.
Key findings
- YOLO-class detectors are the de facto maritime baseline because real-time inference on modest GPU hardware fits shipboard compute budgets. Established
- Published mAP figures on SeaShips-style benchmarks are strong — but the datasets are shore-camera, daylight-weighted, single-geography. Research demonstration
- Small-object recall — kayaks, buoys, jet-skis, containers, persons — is the persistent weakness across every maritime YOLO variant benchmarked. Research demonstration
- Almost no published work reports detector performance under bridge-height pitch/yaw motion, which changes the entire inference geometry. Established
- Performance claims are honest for their test conditions; the marketing extrapolation to 'the ship sees' is not. Maritime-GPT analysis
Why YOLO won the benchmark
Object detection at sea is latency-budgeted: an avoidance decision needs the picture at frame rate, on hardware that fits a console, not a datacentre. Single-pass detectors (YOLO, SSD, RetinaNet) beat two-stage detectors (Faster R-CNN) on that constraint alone. When the dataset literature forked to maritime (SeaShips, SMD), the speed-accuracy trade was already settled in YOLO's favour for the shore-surveillance setting where most data lives. Established
What the scores actually measure
A reported mAP of 80–90% on SeaShips means the model reproduces the dataset's own labels on held-out frames from the same cameras — not that it detects ships in the fog on a moving bridge. The dataset is the test; the test is the dataset. Reporting that number as 'ship detection solved' is a category error the maritime literature mostly avoids but the marketing rarely does. Maritime-GPT analysis
| Regime | Published YOLO evidence | Honest verdict |
|---|---|---|
| Shore/surveillance, daylight | High mAP on SeaShips/SMD | Works — the proven regime |
| Onboard, good visibility | Thin, on small datasets | Plausible, not demonstrated at scale |
| Small objects at range | Consistently the weakest class | The unsolved frontier |
| Fog / night / glare | Sparse; NIR/IR helps partially | Where the camera quits |
| Moving-platform pitch/yaw | Almost unstudied | The gap the papers skip |
The engineering honesty problem
- <strong>Domain shift is the dominant error term.</strong> A detector trained on Singapore's haze and one coastline's camera angle degrades on a different sea state, a different lens, a different wave texture — the failure isn't graceful. Research demonstration
- <strong>Class imbalance mirrors the sea.</strong> Ships are common in the data; kayaks, windsurfers and flotsam are rare — and rare classes are the collision-relevant ones. Maritime-GPT analysis
- <strong>Annotations are the ceiling.</strong> 'Small fishing boat' and 'debris' labels are inconsistent across datasets; the model inherits the label noise. Established
- <strong>Real-time isn't free.</strong> Every accuracy improvement trades frames-per-second; the marine engineer's 15fps budget is the real scoreboard. Maritime-GPT analysis
Where YOLO belongs in the stack
The correct role is the one the research already demonstrates: a fast first-stage detector feeding a fusion engine. YOLO draws the boxes; radar confirms the range; AIS confirms the identity — or flags a detection that reports nothing. The camera adds what no other sensor can (aspect, class, lights); it is not asked to see alone. That architecture is demonstrable today. A YOLO as the lookout is not. Maritime-GPT analysis
What we know
- Real-time ship detection on shore cameras is a solved-enough problem for an assistant layer.
- YOLO variants are the standard, compute-efficient choice for that role.
- The maritime-specific failure catalogue (fog, small objects, motion, glare) is well documented.
What we don't
- Published, verified performance on commercial-ship bridge footage across sea states.
- Whether larger models or fusion-first architectures beat YOLO in the safety-critical regime.
- The real recall rate at COLREGS-relevant ranges on the rare-class objects that kill.
Being tested
- Lightweight backbones for shipboard inference budgets.
- Attention/transformer hybrids vs pure YOLO.
- Self-supervised pretraining to close the data gap.
Maritime-GPT analysis
YOLO earns its place — as a layer, not a lookout. The honest evaluation is 'does it flag a target the fusion engine can verify?' and the answer is yes, within a domain it knows. The safety-critical error is measuring accuracy against the benchmark the model trained on and shipping that number as the lookout's capability. It's a sensor node in a supervised stack — strong there, dangerous alone. Maritime-GPT analysis
This is a research feature, not operational guidance. Simulation results are reported as simulation; demonstrations are not deployments; company claims are attributed as such. Verify safety-critical decisions against official nautical publications and flag-state requirements.
Sources & evidence
The benchmark most YOLO maritime papers score against — knowing its limits is the whole story.
The survey establishing why maritime detection is harder than the benchmark implies.
The canonical single-pass detector reference — the architecture family the maritime literature tunes.
A newer open-water benchmark that exposes the small-object weakness on real water.