YOLO-pose Workflow#
The YOLO-pose path is one validated BehaviorScope-X pose workflow. It uses one YOLO-pose checkpoint for detection, keypoint estimation, and frozen visual-feature extraction, then trains a temporal classifier on pose-derived and visual streams.
Inputs#
You need:
- full-video annotation export from the GUI,
- a class-name file,
- a YOLO-pose
.ptcheckpoint, - a destination folder for sequence caches and training outputs.
Recommended Stage Order#
1. Full-Video Cache#
Open YOLO-pose > Full-Video Cache.
This stage reads full videos and .annot files, runs YOLO-pose, creates crops and pose-derived features, and writes sliding-window NPZ files plus a sequence_manifest.json.
Recommended settings:
- use the same keypoint layout used to train the YOLO-pose checkpoint,
- keep manifest validation enabled,
- use
skip existingwhen resuming an interrupted cache build, - keep the output root specific to the dataset and split.
2. Train/Val Feature Cache#
Open YOLO-pose > Train/Val Feature Cache.
This stage precomputes frozen visual descriptors from the YOLO-pose backbone. Precomputing features makes classifier training faster and easier to reproduce because the temporal model sees a fixed feature cache.
3. Train + Validation Eval#
Open YOLO-pose > Train + Validation Eval.
This stage trains the temporal classifier. The classifier can use:
- pose geometry streams,
- visual descriptors from the YOLO-pose backbone,
- class weighting,
- temporal smoothing or threshold decoding,
- optional model bundling.
When bundling is enabled, training writes a single .pt file containing both the classifier and the YOLO-pose checkpoint reference needed for inference.
4. Inference#
Open YOLO-pose > Inference for one video or YOLO-pose > Batch for a folder.
Use bundled models when possible. They reduce path mistakes because users do not need to supply a separate classifier, config, and YOLO checkpoint.
5. Ethograms + Bouts#
Open YOLO-pose > Ethograms + Bouts to create ethogram timelines and bout-duration summaries from YOLO-pose temporal prediction CSVs. Use *.smoothed_frames.csv files for frame-exact timelines; window prediction CSVs are converted by probability averaging or label voting.
Outputs#
Typical outputs include:
- sequence NPZ cache,
- feature cache,
- classifier checkpoint,
config.json,- training log,
- confusion matrices,
- behavior CSV,
- ethogram and bout summary CSVs,
- optional keypoint CSV,
- optional annotated review MP4.
Use YOLO-pose > Outputs to inventory these artifacts.
When To Use Legacy Clip Cache#
Use YOLO-pose > Legacy Clip Cache only when your dataset is already organized as class-folder clips. For new full-video behavioral work, use full-video annotations and full-video caches.