Common guillemots in the Baltic Sea studied with video surveillance and object detection: raw data, annotations, model, and model outputs
Hentati-Sundberg, Jonas; Olin, Agnes (2022), Common guillemots in the Baltic Sea studied with video surveillance and object detection: raw data, annotations, model, and model outputs , Dryad, Dataset, https://doi.org/10.5061/dryad.xsj3tx9hx
The data comes from common guillemots studied at Stora Karlsö, Sweden between 2019 and 2021. The common guillemots breed at an artificial cliff, and has been filmed continusly from above over three breeding seasons. Using the video material, a YOLOv5 model has been trained to detect adult birds, chicks and eggs. The dataset contains annotations (bounding boxes) used for training the model, the model itself, and outputs from the model (object detections).
The data can be used and shared freely.
The raw data is videos in .avi format, recorded with a Avtech AVH8516 Network Video Recorder (NVR) connected to a 2 megapixel IP camera (Avtech AVM543P) operating at 25 frames per second (FPS). From the videos, single frames have been picked out and annotated with bounding boxes with three classes: adult bird (0), chick (1) and egg (2). Using these annotations, a YOLOv5 model has been trained. The videos have then been downsampled to 1 FPS and Object detection has been run on the whole material, and the output has been stored in a SQLite database coordinates for each detected object.
The dataset contains:
- Annotations (images and bounding box coordinates). The format is "YOLO Darknet TXT", described in detail here: https://roboflow.com/formats/yolo-darknet-txt
- The YOLOV5 model, 2 files: .cfg and .weights
- Inference from the YOLO model (bounding boxes) in SQlite format (.db)
The videos themselves are several Tb (approx 1 Gb per H of film) and therefore not possible to share at this moment at a reasonable cost. Sample videos (downsampled to 1 FPS) are provided as .avi files.
Python code for training and running Yolov5 is available at: https://github.com/BalticSeabird/SeabirdDetections
R code for post-processing the model outputs including the generation of figures is available at https://github.com/BalticSeabird/ObjectDetectionInferences
Svenska Forskningsrådet Formas, Award: 2021-02639
Vetenskapsrådet, Award: 2021-03892
Marcus och Amalia Wallenbergs minnesfond, Award: 2018-0093