Comparison of markerless multi-animal pose tracking and vmTracking in multi-fish tracking.
<p>(A) Overview of the verification process. The yellow box in the schematic represents the processes using multi-animal tracking tools, while the green box represents the processes using single-animal tracking tools. In this study, the evaluation was conducted on a single video consisting of...
Saved in:
| Main Author: | |
|---|---|
| Other Authors: | |
| Published: |
2025
|
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p>(A) Overview of the verification process. The yellow box in the schematic represents the processes using multi-animal tracking tools, while the green box represents the processes using single-animal tracking tools. In this study, the evaluation was conducted on a single video consisting of 2,259 frames. (B) Image of the tracked school of fish. This image corresponds to raw data from a publicly available data set featuring 10 fish swimming in a rectangular aquarium [<a href="http://www.plosbiology.org/article/info:doi/10.1371/journal.pbio.3003002#pbio.3003002.ref031" target="_blank">31</a>]. The videos used in this study are available at: <a href="https://drive.google.com/drive/folders/1Xn3IY46t7DkG2bntbO6sL3oElRPeQA1d" target="_blank">https://drive.google.com/drive/folders/1Xn3IY46t7DkG2bntbO6sL3oElRPeQA1d</a>. Right panel: enlarged view of the rectangle in the left panel. (C) Schematic of virtual marker creation. Five keypoints per fish were tracked, with 2 designated as virtual markers. (D) Swimming trajectories of 10 fish based on the foremost (head) keypoint for maDLC and virtual marker tracking with single-animal DeepLabCut (vmT-DLC). (E) Methods for evaluating tracking. Based on tracking with equidistant keypoints, variance in distance between adjacent keypoints was used as a basic indicator of accurate tracking (larger variance indicates lower accuracy) (upper row). Cosine similarity (Cos sim) of head direction for 10 fish was calculated using angles formed by the 2 foremost keypoints (bottom row). (F) Example photos from 5 scenes and corresponding values for variance (<i>σ</i><sup>2</sup>) and Cos sim. Red arrows in the photos indicate general movement direction of the school of fish. Scene IV did not show a clear direction of movement as a group (not indicated). (G–I) Median variance for each fish (plot) (G), the number of frames showing outliers in variance for each fish based on the interquartile range method (plot) (H), the number of frames for each fish where the head direction changed by more than 90° between frames (approximately 0.03 s) (plot) (I), with bars and error bars representing mean ± 95% confidence interval calculated from the medians of all 10 s fish (<i>n</i> = 10). Statistical analysis was performed using the Wilcoxon signed-rank test (<i>n</i> = 10), **: <i>p</i> < 0.01. (J) Time series changes in Cos sim. Scenes I-V correspond to Scenes I-V in (F). The data underlying these analyses are provided in <a href="http://www.plosbiology.org/article/info:doi/10.1371/journal.pbio.3003002#pbio.3003002.s020" target="_blank">S1 Data</a>. Additionally, these figures can be reproduced using the MATLAB data and code available in the Zenodo repository (<a href="https://doi.org/10.5281/zenodo.14545410" target="_blank">https://doi.org/10.5281/zenodo.14545410</a>). maDLC, multi-animal DeepLabCut; vmT-DLC, virtual marker tracking with DeepLabCut; vmTracking, virtual marker tracking.</p> |
|---|