GUI Overview#
BehaviorScope-X presents the workflow as one annotation surface followed by model-family tabs. This makes the GUI the single validated workflow surface while keeping the differences between pose backbones explicit. Each model-family tab is a pose-model-flexible route into the same downstream cache, classifier, evaluation, and output structure.
Top-Level Tabs#
Annotate + Clip#
Use this tab to import videos, assign splits, define behaviors, draw bout spans, review boundaries, and export full-video annotations. The legacy clip path remains available for older class-folder datasets, but full-video annotation export is the recommended route for new work.
YOLO-pose#
Use this tab when the pose model is an Ultralytics-compatible YOLO-pose checkpoint:
- Build a full-video sequence cache.
- Precompute YOLO-pose visual features.
- Train and validate a temporal behavior classifier.
- Run single-video or batch inference.
- Create ethogram timelines and bout summaries from prediction CSVs.
- Inspect outputs.
This path includes the most interactive inference surface because the trained classifier can be bundled with the YOLO-pose checkpoint.
MobileNetV3#
Use this tab when the pose model is a compatible MobileNetV3 pose checkpoint:
- Build a full-video sequence cache directly from videos with the MobileNetV3 pose checkpoint.
- Extract MobileNetV3 pose-backbone visual features.
- Train and validate the temporal behavior classifier from the MobileNetV3 feature cache.
- Create ethogram timelines and bout summaries from prediction CSVs.
- Use the staged runners for larger held-out suites, static negative-control baselines, and suite-summary collection.
DeepLabCut-HRNet#
Use this tab when the pose workflow is a DeepLabCut SuperAnimal/HRNet project. The stages cover pose and detector fine-tuning, top-down cache creation, HRNet feature-cache extraction, classifier training, held-out evaluation, model-agnostic ethogram/bout summaries, and output inspection.
Ethograms + Bouts#
Each temporal pose-model workflow has an Ethograms + Bouts tab. The tab accepts prediction CSVs from any trained BehaviorScope-X temporal classifier and writes ethogram_segments.csv, ethogram_behavior_summary.csv, and ethogram_summary.json. Use *.smoothed_frames.csv files when you need frame-exact timelines; window-level prediction CSVs are converted by probability averaging or label voting.
Plan Versus Run#
MobileNetV3 and DeepLabCut-HRNet stages have an Action selector:
planprints or writes the commands that would be executed.runstarts the stage.
Use plan before long GPU jobs, especially cache building and classifier training.
Stop Buttons#
- Stop after step asks the process to terminate cleanly when supported.
- Force stop terminates the process immediately. Use it only when the job is stuck or a clean stop is not possible.
Outputs Tabs#
Each model-family tab has an Outputs panel. Use it to inspect sequence caches, feature caches, checkpoints, training logs, confusion matrices, held-out metrics, ethogram exports, bout summaries, keypoint exports, and behavior prediction files.