Skip to content

EDA Tool Plugin

The EDA Tool is an optional workspace for exploring geometric pose features, comparing labeled behavior profiles, and inspecting clusters alongside a matching video. Enable it from Plugins → Manage Plugins..., then select Plugins → Launch EDA Tool.

Requirements and inputs

  • Install the complete user profile (pip install ".[dev,plugins]").
  • Provide frame-indexed YOLO pose-label files from one recording, such as trial_frame_000000.txt, and the ordered keypoint names used during inference.
  • An optional data.yaml supplies keypoint configuration and class names.
  • For video inspection, select the original recording matching those frame indices.

The loader expects whitespace-separated YOLO pose rows. A .csv filename alone does not make a general CSV table, bout summary, or exported feature table a compatible input. Select the inference-label folder explicitly; the tool does not automatically import the current IntegraPose project.

Workflow

  1. In Load Data & Config, select the labels and optional YAML file. Check keypoint order, visibility threshold, and coordinate normalization, then load and preprocess the data.
  2. In Feature Engineering, define skeleton connections or enable all geometric features. Select and calculate the distances, angles, and other features needed.
  3. In Analysis & Clustering, select features and choose individual detections or average behavior profiles. Run optional PCA, followed by hierarchical clustering (AHC) or KMeans.
  4. For AHC on individual detections, a flat-cluster count of zero produces only a dendrogram. Set a positive count to request assignments. AHC on average behavior profiles produces a dendrogram of the behavior means.
  5. Review plots and status text in Visualizations & Output. In Video & Cluster Sync, load the matching video and use playback or the frame slider to highlight the corresponding observations on the feature map.
  6. Use Save Plot to export a figure and Export Data → Export All to save data tables, cluster assignments when available, and supporting metadata.

Interpretation and limits

  • PCA and clustering use standardized selected features. Rows missing any selected feature are excluded from the fit; their assignments remain missing in exports.
  • Clusters describe similarity in those features. They are not automatically validated behaviors, phenotypes, or independent experimental replicates.
  • Class names come from the label configuration. If the pose model labels animals rather than behaviors, its classes must not be interpreted as behavior categories.
  • The video map follows playback and slider position; scatter-point clicking does not seek the video. Use one recording at a time to keep frame identities unambiguous.
  • Behavioral Analytics provides descriptive summaries and a separate advanced bout-analysis workflow. The basic switch count is global across ordered detections, not an animal-specific transition estimate.

Keep the input labels, configuration, selected feature list, and exported results with the study record so the analysis can be reproduced.