UI Components Reference (PoseKit)

This reference describes PoseKit UI components in clean workflow order.

How To Use This Page

• Configure project setup once, then iterate primarily in the main workspace.
• Use model-assist tools only after annotation conventions are stable.
• Keep train/eval/active-learning runs versioned by output folder and seed.

Main Workspace Layout

Area	Role	Key features
Left pane (frame management)	Curate labeling workload.	Labeling frames list, all frames list, search, sorting, random add, smart select, delete selected.
Center pane (canvas)	Perform keypoint annotation.	Image canvas, zoom, drag, keypoint placement, prediction overlays, metadata tags/notes.
Right pane (tool groups)	Configure annotation, display, navigation, model actions.	Annotation controls, display tuning, navigation shortcuts, training/evaluation/active-learning entry points.
Status bar	Runtime feedback.	Save state, progress bar, SLEAP status.

Project Setup Dialog (Wizard / Project Settings)

Purpose

Define project paths, class/keypoint schema, skeleton, and migration behavior.

Controls

Control	Role	Value selection guidance	Common failure mode
Output root	Project artifact root.	Use per-project folder for reproducibility.	Mixing multiple projects in one root.
Labels directory	YOLO pose label storage location.	Keep under output root unless integrating external datasets.	Label path drift between sessions.
Autosave option	Save-on-navigation behavior.	Enable for most workflows.	Disabled autosave causing accidental data loss.
BBox pad fraction	Bounding box expansion from keypoints.	Increase only if extremities clip.	Excessive padding reducing localization precision.
Classes list	Category definitions.	Keep minimal and stable unless multi-class task is required.	Mid-project class changes without migration plan.
Keypoints/skeleton editor	Anatomical schema definition.	Finalize early; version changes deliberately.	Frequent schema churn invalidating older labels.
Migration controls	Existing-label remapping strategy.	Name-based if names are stable; index-based for append-only changes.	Wrong mapping corrupting historical labels.

Annotation Group

Purpose

Control labeling semantics and progression behavior.

Control	Role	How to choose values	Common failure mode
Class selector	Assigns class for current frame.	Keep consistent for single-class projects.	Unintended class flips during rapid labeling.
Keypoint list	Active keypoint selection and status.	Follow canonical anatomical order.	Skipped landmarks due to order confusion.
Mode (frame/keypoint)	Navigation model for labeling.	Frame mode for speed; keypoint mode for consistency checks.	Mode mismatch slowing workflow.
Click visibility	Default visibility code for placement.	Use occluded only when landmark is inferable; missing otherwise.	Visibility misuse reducing training signal quality.

Display Group

Purpose

Tune visualization clarity without changing underlying labels.

Control	Role	How to choose values	Common failure mode
Enhance contrast + settings	CLAHE/sharpening for visibility.	Use conservatively; confirm no hallucinated structure.	Over-enhancement masking true image quality issues.
Show predictions / confidence	Overlay model output for assisted labeling.	Enable during correction passes.	Trusting predictions without manual QA.
Autosave delay	Delayed write frequency.	Keep short enough for crash safety.	Delay too long increases potential data loss window.
Keypoint/edge opacity	Overlay visibility tuning.	Lower when dense overlays hide anatomy.	Overly faint overlays causing placement errors.
Point/text size	Annotation readability scaling.	Increase for high-res imagery or dense skeletons.	Oversized labels obscuring landmarks.
Fit to view	Reset viewport to full frame.	Use after heavy zoom/pan edits.	Remaining zoomed and missing out-of-view mistakes.

Navigation Group

Purpose

Accelerate frame/keypoint traversal and save cadence.

Control	Role	Notes
Prev/Next frame	Sequential frame traversal.	Works with A/D hotkeys.
Save	Immediate save.	Use before long operations.
Next unlabeled	Jump to next incomplete frame.	Critical for completion sweeps.

Model Group

Purpose

Run model-assisted annotation, training, and evaluation workflows.

Subsection	Includes	Value-selection guidance	Common failure mode
Backend selection	YOLO vs SLEAP backend.	Match backend to model family.	Backend/model mismatch.
Prediction controls	Min confidence, current-frame predict, dataset predict, apply predictions, cache clear.	Use higher confidence for conservative adoption.	Bulk apply without review in difficult frames.
YOLO model controls	Weights path, browse/use latest.	Pin known-good weights per project version.	Using stale or incompatible weights.
SLEAP service controls	Conda env, model dir, device, start/stop service.	Validate env/model before long runs.	Service startup issues from bad env resolution.
Training/eval/active learning launches	Dialog entry points.	Use after baseline labels and split quality are confirmed.	Running training from inconsistent labels/splits.

Project Group

Purpose

Manage schema and export artifacts.

Control	Role	Notes
Skeleton editor	Update keypoint topology.	Prefer infrequent, versioned schema changes.
Project settings	Reopen setup/migration controls.	Use for controlled updates only.
Export dataset + splits	Prepare training-ready data layout.	Validate split manifests before training.

Smart Select Dialog

Purpose

Select diverse, high-value unlabeled frames with embedding and clustering support.

Control group	Includes	Guidance
Scope/filtering	Scope, exclusion toggles	Keep scope explicit to avoid hidden frame subsets.
Embedding config	Model, device, batch, max side, enhancement	Tune for stable embeddings first, speed second.
Selection config	N frames, clusters, min/cluster, strategy, threshold	Use cluster coverage to avoid redundancy.

Dataset Split Dialog

Purpose

Create reproducible train/val/test (or k-fold) splits with cluster-aware options.

Control	Guidance
Train/val/test fractions	Maintain enough validation/test signal for reliable comparisons.
K-fold count	Use when dataset size is limited and repeated validation is needed.
Min per cluster	Protect minority clusters from being dropped.
Random seed	Fix seed for reproducible experiment comparisons.
Split name	Use semantic names tied to experiment versions.

Training Runner Dialog

Purpose

Launch and monitor model training/fine-tuning jobs.

Control group	Includes	Guidance
Backend/model	Backend, base weights/model dir	Keep backend/model family aligned.
Training hyperparameters	Batch size, auto-batch, epochs, patience, image size, device	Increase complexity only after baseline convergence is stable.
Data controls	Train fraction, seed, ignore occluded, auxiliary datasets	Keep split/seed fixed for fair model comparisons.
SLEAP export options	Env, output .slp, include aux, embed media	Validate export format before training job launch.

Evaluation Dashboard

Purpose

Compare predictions against labels with keypoint-level and frame-level diagnostics.

Control	Role	Guidance
Backend + model path	Evaluation inference source.	Lock to specific model version for reproducible reports.
PCK and OKS thresholds	Metric sensitivity settings.	Use consistent thresholds across model comparisons.
Output directory	Result artifacts location.	Version by date/run id to track history.
Worst-frame tables	Error triage list.	Feed difficult frames back into labeling queue.

Active Learning Dialog

Purpose

Suggest high-value frames based on uncertainty/disagreement/error signals.

Control group	Includes	Guidance
Strategy	Uncertainty / disagreement / error-focused methods	Start with one strategy, compare lift before combining.
Backend/inference config	Device, image size, confidence, batch, cache	Keep inference settings aligned with evaluation settings.
Scope + N suggestions	Candidate pool and output count	Keep suggestion batches small and reviewable.
Evaluation CSV / keypoint focus	Error-targeted sampling	Use when correcting specific keypoint failure modes.

Recommended PoseKit Workflow

1. Setup project schema and paths.
2. Label pilot subset and validate consistency.
3. Run smart selection to expand diverse coverage.
4. Train baseline and evaluate worst frames.
5. Iterate with active learning and periodic schema-safe exports.