Research map / Video / Video segmentation and tracking
DanceFix: An Exploration in Group Dance Neatness Assessment Through Fixing Abnormal Challenges of Human Pose
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DanceFix, which corrects abnormal 2D skeleton points via bidirectional optical-flow propagation guided by part-level tracking, and introduces the Group Dance Neatness Assessment task with a synthetic fully-neat dataset (DNV) and five metrics.
Problem
Pose estimation from video is unreliable under occlusion and motion blur, which undermines automatic assessment of group dance neatness; no quantitative neatness labels or criteria exist for group dance.
Approach
Combine optical flow (pixel-level motion) with lightweight skeleton data in a bidirectional forward/backward propagation scheme to correct abnormal keypoints, with a part-level tracker (PETD) selecting reliable reference frames; define neatness via limb and joint similarity across dancers.
Key results
- On DNV, DanceFix improves whole-body NS for AlphaPose from 85.74 to 94.25 and average NS gain across eight pose estimators is about 3.54%.
- On JHMDB, DanceFix improves PCK@0.2 by 8.4% for CPM and 9.3% for DeciWatch-style and similar methods across six baselines.
- Ablation on DNV real-world subset shows BOFC+PETD raises AlphaPose whole-body NS from 85.74 to 94.25.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DanceFix pose correction | Existing pose estimators · DNV | NS (%): average increase of 3.54% | Helped |
| Apply DanceFix pose correction | Existing pose estimators · DNV | R-MPJAE: 2.97 | Helped |
| Apply DanceFix pose correction | Existing pose estimators · DNV | R-MPJPE (×100): 0.97 | Helped |
| Apply DanceFix pose correction | Existing pose estimators · DNV | R-MPJAVE: 2.90 | Helped |
| Apply DanceFix pose correction | Existing pose estimators · JHMDB | PCK: 8.4% / 9.3%, 5.9% / 8.9%, 6.0% / 7.3%, 9.3% / 8.7%, 1.4% / 4.7%, and 2.5% / 6.4% | Helped |
| Use BOFC instead of temporal-frame replacement or movement-speed correction (ablation) | DanceFix with AlphaPose · DNV real-world | NS (%): 91.48; 91.88 -> 93.19 | Helped |
| Add PETD to BOFC (ablation) | DanceFix with AlphaPose · DNV real-world | NS (%): 93.19 -> 94.25 | Helped |
| Add PETD to temporal-frame replacement (ablation) | AlphaPose correction · DNV real-world | NS (%): 91.48 -> 92.57 | Helped |
| Add PETD to movement-speed correction (ablation) | AlphaPose correction · DNV real-world | NS (%): 91.88 -> 92.22 | Helped |
| Add PETD to BOFC (ablation) | DanceFix with AlphaPose · DNV virtual | NS (%): 95.02 -> 95.02 | No clear effect |
| Add PETD to temporal-frame replacement (ablation) | AlphaPose correction · DNV virtual | NS (%): 94.97 -> 95.00 | No clear effect |
| Add PETD to movement-speed correction (ablation) | AlphaPose correction · DNV virtual | NS (%): 94.89 -> 94.92 | No clear effect |
Surprising findings
- Abnormal skeleton data often persists across multiple consecutive frames in dance, so checking only adjacent frames is insufficient.
Limitations
- Assumes all dancers follow the same choreography.
- Operates on 2D poses only; 3D extension left to future work.
- Real-world videos sourced from video websites; consent from creators still being sought.
Experimental setup
Datasets: Dancer Parts JHMDB DNV
Benchmarks: DNV JHMDB
Metrics: NS R-MPJAE R-MPJPE R-MPJAVE R-MPJPVE PCK@0.2 PCK@0.1
Baselines: AlphaPose HRNet RSN LiteHRNet PVT PVT v2 ED-Pose DGN SHaRPose CPM SBL PoseWarper
Builds on: ByteTrack FlowFormer DeciWatch
Closest papers by meaning
- DanceCamera3D: 3D Camera Movement Synthesis with Music and Dance (CVPR 2024 · similarity 0.84)
- DanceMVP: Self-Supervised Learning for Multi-Task Primitive-Based Dance Performance Assessment via Transformer Text Prompting (AAAI 2024 · similarity 0.84)
- DisCo: Disentangled Control for Realistic Human Dance Generation (CVPR 2024 · similarity 0.83)
- Harmony4D: A Video Dataset for In-The-Wild Close Human Interactions (NeurIPS 2024 · similarity 0.83)
- HardMo: A Large-Scale Hardcase Dataset for Motion Capture (CVPR 2024 · similarity 0.83)
- DanceAnyWay: Synthesizing Beat-Guided 3D Dances with Randomized Temporal Contrastive Learning (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.