Post-Hoc Analysis Guide
You've trained a model and run inference to get keypoint and behavior data. Now what? This guide introduces advanced post-hoc analyses to help you extract deeper scientific insights from your raw tracking data and uncover subtle behavioral phenotypes.
HMM-VAE-LSTM Video Segmentation
This unsupervised approach discovers the underlying structure of behavior without pre-defined labels. It learns a compressed representation of postures (VAE), models their temporal dynamics (LSTM), and segments them into discrete, stereotyped behavioral syllables (HMM).
Use this for:
- Unsupervised discovery of behavioral motifs.
- Analyzing the syntax and sequencing of actions.
- Identifying novel behavioral states affected by your experiment.
Sub-Behavior Discovery with LSTM Autoencoders
This Seq2Seq model learns a compressed "signature" for entire behavioral sequences. By clustering these signatures, it automatically discovers stereotyped, repeatable variations within a single behavior class (e.g., fast vs. slow grooming).
Use this for:
- Finding fine-grained motor patterns within a broad behavior.
- Quantifying an animal's repertoire of specific actions.
- Identifying subtle motor phenotypes that differ between groups.
LSTM-Based Sequence Classifiers
Go beyond single-frame classification. An LSTM classifier considers a sequence of keypoints over time to make a prediction. This allows it to learn dynamic patterns that are invisible to a standard frame-by-frame model, capturing the *how* of a movement, not just the *what*.
Use this for:
- Classifying behaviors defined by their dynamics (e.g., hesitant vs. confident walk).
- Distinguishing between similar postures with different meanings.
- Improving classification accuracy in noisy conditions.
Advanced Gait & Kinematic Analysis
Move beyond simple speed calculations. This involves extracting detailed kinematic variables from limb keypoints during locomotion to build a comprehensive profile of an animal's gait. Analyze stride parameters, inter-limb coordination, and postural dynamics.
Use this for:
- Detailed analysis of locomotor deficits in disease models.
- Quantifying changes in balance and coordination.
- Comparing gait signatures between different experimental groups.