Pole-Arina: A Privacy-Preserving Dataset and Benchmark for Static Pole Tricks

Diana Marin, Katharina Scheucher, Peter Kán
Technische Universität Wien, Austria
Paper Dataset Code
Pole-Arina teaser showing skeletons and dense optical flow.

Pole-Arina replaces identifiable RGB with privacy-preserving motion representations. We extract 2D pose skeletons and dense optical flow, enabling clip-level trick recognition, frame-wise phase understanding, and coaching-oriented analysis for static pole tricks under strong privacy constraints.

Abstract

Computer vision for sports lacks datasets and models that target pole sports, an activity marked by self-occlusion, body inversions, and sustained contact with the pole. We address this gap by introducing Pole-Arina, a curated, privacy-preserving dataset for markerless analysis of static pole tricks.

The collection comprises 836 videos from 58 participants and provides two appearance-suppressing modalities derived from the recordings: pose skeleton sequences (2D joints) and dense optical flow. We annotate each clip with both per-video trick labels and per-frame labels that capture temporal structure, enabling evaluation of clip-level and frame-wise recognition under privacy constraints.

As reference baselines, we benchmark lightweight temporal models across all benchmark settings and analyze common confusions and imbalance effects. In addition, we provide a geometry-aware analysis module that measures trick-specific body orientation, joint alignments, and proximities to produce interpretable overlays and actionable feedback.

BibTeX

@inproceedings{polearina2026,
  title     = {Pole-Arina: A Privacy-Preserving Dataset and Benchmark for Static Pole Tricks},
  author    = {Diana Marin and Katharina Scheucher and Peter Kán},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  month     = {June},
  year      = {2026},
}