Overview of the Fishnet Open Images Database
The Fishnet Open Images Database is a dataset of images from on-board monitoring cameras that have been annotated with object bounding boxes. The current dataset is composed entirely of long line tuna fishing activity in the Western and Central Pacific. In the future we anticipate contributions from other fisheries that could benefit from having AI algorithms improve their workflows.
The v1.0.0 dataset contains 549,209 bounding boxes for 34 object classes (composed entirely of fish species and humans) on 143,818 images collected from 96 unique cameras with an average of 4 bounding boxes per image. These data also include Food and Agriculture Organization (FAO) major fishing area (MFA) for which the image occurs. The MFAs can provide an indication of likely species encountered. These data also have unique ids for cameras, sequences, and frames within sequences. These data have splits for model training, validation, and testing.
The bounding box annotations have been manually drawn by professional annotators at Samasource as well as IKE Solutions reviewing outputs of model inference utilizing the Tator application by CVision. The dateset principally annotates fish, however, there is value in annotating humans as well as they can be a proxy for fishing activity. For fishes, we define a fish feature or part (e.g. fins, tail, head) if it is ≥80% in the frame. We label both fish on the deck of the vessel and fish that are recognizable as fish alongside the vessel in the water. We also annotate bounding boxes around ‘human features’ that are not faces, such as boots, legs, and arms.
In processing both video and declaration data we remove any information related to location, vessel name, or unique identifying features like vessel license number. Human faces are blurred.
For more details, you can read this paper about how we evaluate the performance of existing detection and classification algorithms and demonstrate that this dataset can serve as a challenging benchmark for development of computer vision algorithms in fisheries.
The dataset was compiled by The Nature Conservancy in partnership with Satlink, Archipelago Marine Research, the Pacific Community, the Solomon Islands Ministry of Fisheries and Marine Resources, the Australia Fisheries Management Authority, and the governments of New Caledonia and Palau.