Research map / Video / Action recognition
CPR-Coach: Recognizing Composite Error Actions based on Single-class Training
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.
Builds CPR-Coach, a multi-view video dataset of 13 single CPR errors and 74 composite errors, and proposes ImagineNet, which fuses features of single-error samples to recognize unseen composite errors under single-class training.
Problem
Recognizing composite (multi-error) actions in CPR skill assessment when only single-error classes are available for training ("Single-class Training & Multi-class Testing"), since exhaustively collecting all 8,191 composite combinations is infeasible.
Approach
ImagineNet trains on pairs of single-class video features fused via FC, self-attention, or cross-attention (with random MixUp-style weighted summation) and a BCE multi-hot loss, so that the model learns to recognize combined error features it never saw directly.
Key results
- ImagineNet-FC improves TSM's composite-error mAP from 0.5662 to 0.7053 (+13.91%) and mmit mAP from 0.6618 to 0.7566 (+9.48%) when trained only on single-class Set-1.
- ImagineNet-FC improves MMNet's mAP from 0.5948 to 0.6927 (+9.79%) and mmit mAP from 0.6735 to 0.7478 (+7.43%).
- ImagineNet-CA with RGB+Pose reaches mAP 0.7110 with 0.1642 ms latency, surpassing MMNet (0.6927 mAP) and Two-Stream (0.6003 mAP) under cross-modality settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply ImagineNet-FC to TSM instead of direct migration | TSM on CPR-Coach Set-2 · CPR-Coach Set-2 | mAP: 0.5662 -> 0.7053 | Helped |
| Apply ImagineNet-FC to TSM instead of direct migration | TSM on CPR-Coach Set-2 · CPR-Coach Set-2 | mmit mAP: 0.6618 -> 0.7566 | Helped |
| Replace random weighted-sum aggregation with the vanilla feature aggregation (ablation) | ImagineNet-FC with TSM · CPR-Coach Set-2 | mAP: 0.7053 -> 0.6871 | Hurt |
| Replace random weighted-sum aggregation with CBP (ablation) | ImagineNet-FC with TSM · CPR-Coach Set-2 | mAP: 0.7053 -> 0.6434 | Hurt |
| Replace random weighted-sum aggregation with BLOCK (ablation) | ImagineNet-FC with TSM · CPR-Coach Set-2 | mAP: 0.7053 -> 0.6569 | Hurt |
| Add positional embeddings (ablation) | ImagineNet-SA with TSM · CPR-Coach Set-2 | mAP: 0.6822 -> 0.7011 | Helped |
| Add positional embeddings (ablation) | ImagineNet-CA with TSM · CPR-Coach Set-2 | mAP: 0.6725 -> 0.6752 | Helped |
| Replace self-attention fusion with cross-attention fusion (ablation) | ImagineNet on TSM · CPR-Coach Set-2 | mAP: 0.7011 -> 0.6752 | Hurt |
| Increase self-attention depth from one to two layers (ablation) | ImagineNet-SA with TSM · CPR-Coach Set-2 | mAP: 0.7011 -> 0.7007 | Hurt |
| Increase self-attention depth from two to three layers (ablation) | ImagineNet-SA with TSM · CPR-Coach Set-2 | mAP: 0.7007 -> 0.6995 | Hurt |
| Increase cross-attention depth by adding self-attention (ablation) | ImagineNet-CA with TSM · CPR-Coach Set-2 | mAP: 0.6752 -> 0.6766 | No clear effect |
| Fuse RGB and pose features with ImagineNet-CA rather than Two-Stream | TSM + ST-GCN · CPR-Coach Set-2 | mAP: 0.6003 -> 0.7110 | Helped |
Surprising findings
- PoseC3D performs worse than ST-GCN on single-class CPR recognition, attributed to the repetitive and circular nature of CPR breaking its 3D heatmap representation.
- Direct migration of single-class models to composite errors causes large performance drops (e.g., TSM mAP 0.5662 on Set-2 vs. 0.9901 Top-3 on Set-1 single-class).
Limitations
- Composite error set limited to 59 paired, 10 triple, and 5 quadruple combinations selected by doctors rather than all 8,191 combinations.
- At inference the single composite-error input is duplicated to fill the two-input architecture, a train/test mismatch.
- Deployment evaluation uses only 187 videos from a hospital collection, with preliminary results.
Experimental setup
Datasets: CPR-Coach Set-1 CPR-Coach Set-2
Benchmarks: CPR-Coach Set-1 CPR-Coach Set-2
Metrics: Top-1 accuracy Top-3 accuracy mAP mmit mAP latency GFLOPs
Baselines: TSN TSM I3D TPN C3D TIN SlowFast TimeSFormer ST-GCN PoseC3D Two-Stream MMNet
Builds on: TSN MixUp Two-Stream MMNet
Closest papers by meaning
- Contextualized Policy Recovery: Modeling and Interpreting Medical Decisions with Adaptive Imitation Learning (ICML 2024 · similarity 0.80)
- Modeling Multiple Normal Action Representations for Error Detection in Procedural Tasks (CVPR 2025 · similarity 0.80)
- CholecTrack20: A Multi-Perspective Tracking Dataset for Surgical Tools (CVPR 2025 · similarity 0.79)
- BenchX: A Unified Benchmark Framework for Medical Vision-Language Pretraining on Chest X-Rays (NeurIPS 2024 · similarity 0.79)
- FineParser: A Fine-grained Spatio-temporal Action Parser for Human-centric Action Quality Assessment (CVPR 2024 · similarity 0.79)
- VILA-M3: Enhancing Vision-Language Models with Medical Expert Knowledge (CVPR 2025 · similarity 0.79)
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.