Roadmap
🔜 Q2 2025
- Add support for SLEAP (#8)
- Add support for saving and loading HMMs (#14)
- Speed up HMM fitting by training on a random subset of data (#22)
- Add native support for parallel HMM scanning in CLI (#23)
- Test training LISBET classifiers using HMM prototypes (#24)
- Add native support for n > 2 individuals (#28)
- Enable segmentation mask generation from keypoints (#38
- Testing “multi-dyadic” strategy for inference on groups (n > 2) (#27)
đź“… Q3 2025
- Introduce “Switch Individuals” as a data augmentation strategy
- Conduct ablation study on pretraining tasks
đź”® Future Ideas
- Explore end-to-end models for discovery-driven behavior labeling (#21)
- Add support for raw video input
- Add support for segmentation mask input
- Compare keypoints vs. keypoint-derived segmentation mask performance
- Evaluate frame-based models (raw video and segmentation masks) for n > 2 individuals
- Refactor CLI (#20)
- Introduce GUI
- Explore the use of LLMs for behavior classification from embeddings
- Automate model benchmarking