A new pipeline turns hours of overlooked marina camera footage into a tiny, useful training set for spotting ships automatically.
The method starts by fixing a basic problem with old security-style cameras: they pan, tilt, and zoom, but nobody logged where they were pointing at any given moment. Using two image-matching tools called SuperPoint and LightGlue, researchers matched individual video frames back to a reference panorama to recover that missing position data. Each frame then got tagged with weather and time-of-day conditions, and a sampling step picked out frames that varied in viewpoint and conditions rather than just grabbing duplicates. A second pass specifically hunted for an under-represented case, ships far off near the horizon, by clustering visual patches with a statistical model. The payoff: 40,718 candidate frames shrank to 220 labeled images, which were enough to fine-tune a detector hitting 94.78% accuracy (AP50) in testing at a real facility, the CMMI MDigi-I Smart Marina testbed.
The bottleneck in most computer vision work isn't collecting video, it's the tedious job of drawing boxes around objects in thousands of near-identical frames. This result suggests smarter sampling, not more raw footage, is the cheaper path to a working detector, which matters for the harbors, ports, and marinas that already have idle PTZ cameras but no labeled data to show for it.
One testbed and one camera rig is not proof this generalizes. A 99.5% reduction sounds great until a different harbor's fog, glare, or boat traffic breaks the assumptions baked into this pipeline.