Skip to main content
. 2024 Feb 22;3(2):e0000432. doi: 10.1371/journal.pdig.0000432

Fig 1. Data acquisition and analysis pipeline.

Fig 1

A. Acquisition of 503 videos using the dedicated Baby Moves smartphone app [12]. B. 100 videos were selected for DLC training, stratified by age at video acquisition, sex and GMs classification. C. From each of the 100 training videos, five frames were selected for manual labelling using a k-means clustering algorithm (see Methods; total DLC training dataset: 500 frames). D. The trained DLC model was used to label all frames in all videos. This constitutes the full dataset with body point positional data used for GMs classification E. Labelling accuracy was evaluated in a subset of 50 frames not included in the training dataset. The difference between manual and automatic labelling (Manual/auto) and inter-rater reliability (IRR) of two annotators was calculated and expressed as percentage of infant length.