Maritime dataset 
Until 2023, no sufficiently comprehensive or annotated dataset was published for nautical objects 
ather than ships (e.g. buoys, obstacles, and navigation aids). 
In order to train and evaluate an object detection model, an appropriate dataset is required. Maritime 
datasets are rare because the collection of data is time-consuming and costly. When selecting training 
and validation data, caution should be exercised. It should also be considered how this choice affects 
performance evaluation, especially in comparative assessments. Depending on the purpose of the 
avaluation, the use of different training data can lead to a biased interpretation of the results. In par- 
ticular, any performance gap in such cases should be attributed to the combination of algorithm and 
training data rather than just the algorithm (ISO, 2022). 
Singapore Maritime Dataset: 
The Singapore Maritime Dataset was recorded with a Canon 70D camera in July 2015 and May 2016. 
It contains 40 videos recorded from an on-shore mounted system. Eleven additional videos recorded 
aboard a vessel are part of this dataset. Furthermore, 30 videos in near infra-red are available (Prasad, 
2024). The dataset also provides annotations that can be used for training Al systems for purposes such 
as vessel tracking, traffic analysis, and security measures. In Kim et al. (2022) , the number of classes is 
reduced in order to enhance the training possibilities. 
ABOShip maritime dataset 
The ABOShip Turku dataset was acquired from a set of 135 videos collected from a sightseeing water- 
craft by a RGB daylight camera with a field of view of 65° and stored in HD (1920x720 pixel) resolution 
at 15 frames per second in MPEG format. The route of the watercraft extended from the city of Turku 
to Ruissalo in south-west Finland. The videos comprise the urban area along the Aura river, the port, 
and the Finnish Archipelago and were recorded over 13 days (26 June 2018 — 8 July 2018). This dataset 
nas annotations, which differentiate between the classes seamark, motorboat, sailboat, passenger 
ship, cargo ship, military ship, ferry, misc boat, and miscellaneous (lancu et al., 2021). 
Annotation and selection of appropriate object classes 
If every manufacturer defines and selects their own object classes, the navigational personnel needs 
to be aware of the individual capabilities and limitations of the equipment. A likely conclusion is that 
stressful situations lead to a higher probability of incorrect interpretation of the data displayed. Fol- 
l(owing this assumption, a common object nomenclature is advisable. 
All of the maritime datasets examined contain a classification of “null” to indicate that no object is 
present in this image of the dataset. The power-driven vessels defined in COLREG Rule 3 (b) have been 
found to be clustered as “boat”, “ferry”, “vessel”, “container”, “cruise”, “speedboat”, “passenger 
ship”, “cargo ship”, “military ship”, or “warship”. All samples reviewed used “sailboat” according 
COLREG Rule 3 (c). The terms “seamark” or “buoy” are frequently used in the classification. The field 
study was unable to prove whether object recognition can also correctly detect a “vessel restricted in 
her ability to manoeuvre”, which is defined in COLREG Rule 3 (g) with Rule 27 “vessel constrained by 
her draught” defined in COLREG Rule 3 (h) “detect the sign” defined in COLREG Rule 28. 
Which information should be grouped in one category and presented to the seafarer is to be deter- 
mined by the respective IMO working party. A systematic survey of nautical personnel in order to de- 
termine the needs and preferences of the users of such systems is considered useful.