#### **This OSF component contains features that were used to construct [SEM models](https://psyarxiv.com/pt6hx)**
# Features before resampling to 3Hz
This directory contains raw features before resampling for each activity (video). Those features are: object appearance/disappearance (binary), optical flow features (continuous), names and weighted average (1/distance^2) semantic embeddings of 3 objects nearest to the right hand of the actor, average semantic embeddings of all objects in the current frame, pose and motion features.
# Resampled Data: before and after PCA
This directory contains resampled data (e.g. *1.1.1_kinect_sep_09_all_features_resampled.csv*): roughly, the processing script perform horizontally concatenating all features, dropping rows that one of the features have NAs, and then resampling to 3Hz. [Code](https://github.com/mbezdek/extended-event-modeling/blob/main/src/preprocess_features/preprocess_indv_run.py#L220C12-L220C12)
This directory also contains PCA-ed resampled data (e.g. *1.1.1_kinect_sep_09_all_pcs_input_vector.csv*): resampled data contains 253 features in total (2 object appearance/disappearance, 2 optical flow info, 149 pose and motion info, 100 nearest object and all object embeddings). They are projected to a 30 dimensional space. Whitening PCA was applied independently for each type of features (hence 4 PCA component matrices and 4 means). [Code](https://github.com/mbezdek/extended-event-modeling/blob/main/src/train_eval_inference/run_sem_pretrain.py#L236)
# PCA components and means
This directory contain principal components (derived from all videos) and means (derived from all videos). These components and means are for storage purpose, PCA-ed data are in **Resampled Data: before and after PCA**
# Data Loader
If you want a quick way to load the PCA data, along with hierarchical event labels (e.g. exercise -> cardio_exercise -> jump rope), please check out this repository: https://github.com/qihongl/META-loader. There is a single Python class called METAVideos, and it has functions to:
- Load all video data with labels;
- Load a single video given the video ID;
- Get the RDM over event types;
- Transition structure over event types; etc.
