DIGI4ECO Underwater Fish Tracking Dataset

Résumé fourni par la source

Description This dataset contains manual multi-object tracking (MOT) annotations for four underwater video sequences acquired at three marine observatory sites within the DIGI4ECO Horizon Europe project (Grant Agreement No. 101112883). The dataset was produced to support the development and evaluation of automatic fish tracking algorithms for fixed and mobile underwater monitoring platforms in temperate and coastal marine environments. All annotations were produced manually using the CVAT annotation platform and exported in CVAT XML format. Each annotation file follows the standard CVAT 1.1 schema and includes, for each tracked individual: bounding box coordinates (xtl, ytl, xbr, ybr) in pixel space, a per-frame occlusion flag, and an outside flag indicating when the individual is not visible in the frame. All individuals are annotated under the single class Fish, as species-level identification from bounding box annotations alone is not reliable in these sequences. Dataset contents The dataset comprises four annotated video sequences: Ancona (Adriatic Sea, Italy). One sequence of 1,200 frames acquired from an Adriatic Sea underwater video survey at 1920×1080 pixels. The sequence contains 86 individually tracked fish in a high-density school scenario with a 17.0% occlusion rate and near-constant bounding box overlaps (93% of frames), representative of the conditions expected at the Ancona DIGI4ECO demo site. OBSEA (Mediterranean Sea, Spain). One sequence of 150 frames acquired by the OBSEA seafloor cabled observatory at 20 m depth, 4 km off the coast of Vilanova i la Geltrú, Barcelona, at a resolution of 1912×1080 pixels. The sequence contains 12 individually tracked fish under moderate-density, low-occlusion conditions typical of a Mediterranean artificial reef community. SmartBay Dataset 1 (Galway Bay, Ireland). One sequence of 2,400 frames acquired by the ANERIS-EMUAS1 camera at the SmartBay Observatory at 1920×1080 pixels. The sequence contains 92 individually tracked fish and presents the most challenging tracking conditions of the four sequences: poor visual contrast between fish and the sandy substrate due to light reflection and camera colour response, indistinct fish contours, erratic movement, and near-universal bounding box overlaps (92% of frames). Four tracks include re-entry events, where a fish temporarily becomes indistinguishable from the background and reappears. SmartBay Dataset 2 (Galway Bay, Ireland). One sequence of 2,101 frames acquired by SmartBayCam1 at the SmartBay Observatory at 1920×1080 pixels. The sequence contains 301 individually tracked fish in an open-water schooling scenario with high fish density (average 26.2 fish per frame) and good visibility, providing a complementary high-density benchmark with dynamic schooling patterns and frequent entry/exit events. Summary statistics Sequence Frames Tracks Avg fish/frame Occlusion rate Frames with sig. overlap Ancona 1,200 86 22.5 17.0% 93% OBSEA 150 12 9.2 2.8% 25% SmartBay DS1 2,400 92 17.8 7.0% 92% SmartBay DS2 2,101 301 26.2 7.7% 88% Total 5,851 491 File organisation DIGI4ECO_tracking_dataset/├── README.md├── Ancona/ │ ├── video.mp4 │ └── annotations.xml ├── OBSEA/ │ ├── video.mp4 │ └── annotations.xml ├── SmartBay1/ │ ├── video.mp4 │ └── annotations.xml ├── SmartBay2/ | ├── video.mp4 | └── annotations.xml└── figures/ ├── Ancona.png ├── OBSEA.png ├── SmartBay1.png └── SmartBay2.png Annotation format Annotations are provided in CVAT XML format (version 1.1), compatible with the CVAT annotation platform (https://github.com/opencv/cvat). Each XML file contains a block with task metadata and label definitions, followed by a series of elements. Each track has a unique integer ID and contains one element per frame, with pixel-space bounding box coordinates, an outside flag (1 = individual not visible in frame), and an occluded flag (1 = individual partially or fully hidden). Tracks are defined over contiguous frame ranges and include outside boxes at the end of the track to indicate its termination. Related resources DIGI4ECO project: https://digi4eco.eu CVAT annotation tool: https://github.com/opencv/cvat OBSEA ERDDAP server: https://data.obsea.es/erddap/index.html SmartBay ERDDAP server: https://erddap.marine.ie/erddap/index.html

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.