Kinetic Mining in Context: Few-Shot Action Synthesis via Text-to-Motion Distillation

28th International Conference on Pattern Recognition (ICPR 2026) — main conference proceedings

---

KineMIC adapts a pre-trained Text-to-Motion diffusion model into a specialized Action-to-Motion generator for Human Activity Recognition (HAR), using as few as 10 real samples per class. It leverages CLIP semantic correspondences to mine kinematically relevant motion from a large source dataset, guiding fine-tuning of the generalist backbone via contrastive distillation and LoRA adaptation.

For method details, results, and animated examples → project page · paper

Visual Comparison

MDM (baseline)	KineMIC (ours)

Repository Structure

KineMIC/
├── external/
│   ├── motion-diffusion-model/   # adapted MDM — training & sampling scripts
│   └── pyskl/                    # ST-GCN evaluator for downstream HAR
├── scripts/                      # data preprocessing & few-shot split tools
├── data/                         # datasets (NTU60, NTU120, HumanML3D, ...)
├── prep/                         # setup shell scripts
└── docs/
    ├── setup.md                  # environment & data setup
    ├── train.md                  # training reference
    ├── sample.md                 # sampling reference
    └── other.md                  # misc utilities & tools

Getting Started

Step	Guide	Description
1	Setup	Environment, dependencies, data download
2	Training	MDM baseline, KineMIC, ST-GCN evaluator
3	Sampling	Motion synthesis, synthetic dataset generation
—	Other	Few-shot split tools, ST-GCN evaluator, misc utilities

Citation

If you find this work useful, please cite:

@misc{cazzola2026kineticminingcontextfewshot,
      title={Kinetic Mining in Context: Few-Shot Action Synthesis via Text-to-Motion Distillation}, 
      author={Luca Cazzola and Ahed Alboody},
      year={2026},
      eprint={2512.11654},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.11654}, 
}