Justification and tuning of the change-point detection (CPD) method.
<p>(A) Video-based, frame-level human annotations (colored areas) and detected change points (vertical gray lines) overlaid on IMU data from the 14-second scratching episode shown in <a href="http://www.plosbiology.org/article/info:doi/10.1371/journal.pbio.3003431#pbio.3003431.s024&quo...
Збережено в:
| Автор: | |
|---|---|
| Інші автори: | , , , , |
| Опубліковано: |
2025
|
| Предмети: | |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Резюме: | <p>(A) Video-based, frame-level human annotations (colored areas) and detected change points (vertical gray lines) overlaid on IMU data from the 14-second scratching episode shown in <a href="http://www.plosbiology.org/article/info:doi/10.1371/journal.pbio.3003431#pbio.3003431.s024" target="_blank">S1 Movie</a>. Lowercase letters indicate the following behaviors: (a) sniffing with head pitched downward, (b) left body scratching, (c) left hind paw licking, (d) head scratching on the left side, and (e) decelerating head scratching on the left side. Change points were detected using Gaussian kernel-based Pruned Exact Linear Time (PELT) CPD applied to z-scored angular velocity (<b><i>ω</i></b>) and gravitational acceleration () data. Total acceleration (<b>a</b>) is shown alongside in lighter shades. (B) Number of detected change points (green) and median segment duration (purple) as a function of the penalty parameter (<i>λ</i>), computed from z-scored and <b><i>ω</i></b> data in a 30-minute example recording. (C) Adjusted Rand Index (ARI) quantifying the similarity between segmentations obtained by running CPD on the full 30-minute recording versus running CPD in parallel on non-overlapping fixed-length windows. Notably, parallel CPD, which is computationally faster, produces nearly identical segmentations to the full-recording approach when window sizes are at least 5 minutes. (D) Histogram of mean angular speed values (i.e., the norm of gyroscope measurements) computed over non-overlapping, fixed-length windows of 700 ms. The x-axis is logarithmic to highlight the bimodal distribution, which reflects periods of immobility and movement. A threshold of 12 °/s on the standard deviation of angular speed was applied to isolate immobility periods. (E) Average log-likelihood (LL) computed across CPD segments or fixed-length windows from 10 independent GMM fits, plotted as a function of the number of mixture components <i>K</i>. LL values are shown with and without the ablation of immobility segments or windows prior to modeling (see main text for more details). For each <i>K</i>, the shaded region spans the full range of LL values across the 10 fits. Solid lines connect the maximum LL values at each <i>K</i>, indicating the best-fitting model.</p> <p>(TIFF)</p> |
|---|