End-to-End Workflow
TrackerKit Workflow
- Load video and outputs
- Set input video, CSV output, optional rendered video output.
- Calibrate detection
- Pick detection mode.
- Use preview/test detection before full run.
- Configure tracking
- Set
MAX_TARGETS, assignment distance, track lifecycle thresholds.
- Run forward/backward tracking
- Enable backward pass for better conflict resolution.
- Post-process and export
- Resolve identities, interpolate gaps as needed.
- Save final CSV and optional diagnostics.
PoseKit Workflow
- Load image set and project settings.
- Label or refine keypoints frame-by-frame.
- Use tools (smart select, metadata tags, split generation).
- Export/prepare training-ready datasets.
Decision Points That Matter Most
- Detection mode affects raw input quality to tracker.
- Reference body size scales multiple heuristics.
- Post-processing can fix or amplify detection mistakes depending on thresholds.
Failure Pattern Checklist
- If targets merge often: tighten morphology and assignment distance.
- If tracks fragment: increase recovery/lost frame thresholds and validate detection confidence.
- If runtime is slow: reduce resize factor, disable non-critical overlays/histograms, verify GPU backend.
ClassKit Workflow
- Create or open a project with source image directories.
- Ingest and embed crops using a backbone model.
- Cluster embeddings and visualize with UMAP.
- Label identity classes manually or via AprilTag auto-labeling.
- Train a classification head and evaluate results.
- Export labeled datasets for downstream use.
DetectKit Workflow
- Curate detection training datasets from TrackerKit exports.
- Configure YOLO training parameters.
- Launch training and monitor loss curves.
- Evaluate model performance on validation sets.
FilterKit Workflow
- Load a dataset directory.
- Apply filtering criteria (quality, diversity, metadata).
- Export the filtered subset for training or analysis.
RefineKit Workflow
- Load tracked trajectories and source video.
- Review flagged suspicious segments in the suspicion queue.
- Correct identity assignments using the interactive canvas.
- Export refined trajectories.