CURVATURE-GUIDED TASK SYNERGY FOR SKELETON BASED TEMPORAL ACTION SEGMENTATION (ICLR 2026)

Status: Code is available, pretrained models coming soon.
Paper: Curvature-Guided Task Synergy for Skeleton-based Temporal Action Segmentation (ICLR 2026)
ArXiv/OpenReview: coming soon

---

TL;DR

CurvSeg restores cross-task synergy between action classification and boundary localization for skeleton-based temporal action segmentation by:

1. using representation curvature as a geometric prior for boundary detection, and
2. using localization to refine classification feature geometry via a bidirectional consistency objective,
   plus an Expert-Driven Decoupling (EDD) module (Mixture-of-Experts) to provide task-adaptive features.

---

Results

Quantitative Results

• Main tables / comparisons: coming soon

Qualitative Results

• Visualization examples: coming soon

---

Installation

Environment

• PyTorch 1.10.1 (required)

Setup

1. Install PyTorch 1.10.1 (match your CUDA version).
2. Install remaining dependencies:

pip install -r requirements.txt

Datasets

We evaluate on:

• MCFS-22 / MCFS-130
• PKU-MMD
• LARa

Links: Google Driver

Usage

We provide one unified script for training + testing. For different datasets, we provide separate .sh launchers.

Quick Start

# Example (placeholder)
bash pku_sub.sh

Checkpoints

Pretrained models: coming soon

Citation

If you find this work helpful, please cite:

@inproceedings{curvseg2026,
  title     = {Curvature-Guided Task Synergy for Skeleton-based Temporal Action Segmentation},
  author    = {Coming soon},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026}
}

License

This repository is released under the coming soon license. (Recommendation: choose a standard OSI-approved license before release.)

Acknowledgements

This codebase builds upon / refers to:

DeST (coming soon)

LaSA (coming soon)

Other dependencies: coming soon