Skip to main content

Table 2 Dataset parameters and accuracy metrics

From: High-resolution, non-invasive animal tracking and reconstruction of local environment in aquatic ecosystems

DatasetAnnotationsRate (Hz)Resolution (px)Coverage (%)Accuracy metrics
     MetricReconstruction (cm)Reprojection (px)Tracking (cm)
single171302.7k97.79median0.309.65NA
     RMSE1.2816.30NA
w/ svas above100.00as above
mixed80304k69.60median0.443.77NA
     RMSE1.097.77NA
school160602.7k78.38median0.062.57NA
     RMSE0.303.78NA
w/ svas above94.02as above
accuracy73304k80.64 ±16.73median-0.14 ±0.063.53 ±1.960.14 ±0.33
     RMSE1.34 ±0.798.56 ±5.211.09 ±0.47
w/ svas above97.29 ±2.20medianas above0.28 ±0.32
     RMSE  2.12 ±1.37
  1. w/ sv’ indicates that trajectory points were also estimated from single-view projections at an interpolated depth component. Annotations lists how many frames were annotated for training Mask R-CNN, Rate the frames per second of each video set, i.e. the temporal tracking resolution. Resolution is video resolution, 2.7k: 2704 ×1520 px, 4k: 3840 ×2160 px. Coverage is the mean coverage off all individual trajectories of a dataset. Reconstruction metrics refer to the deviation of reconstructed camera-to-camera distances from the actual distance, Reprojection metrics to the reprojection of triangulated 3D tracks to the original video pixel coordinates and Tracking to the deviation of the tracked calibration wand length from its actual length. In case of the ’accuracy’ dataset, the accuracy results are listed as the mean and standard deviation of the four repeated trials. NA: not applicable