Research map / Label-efficient and robust learning / Recognition and long-tail
Prototypical Calibrating Ambiguous Samples for Micro-Action Recognition
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 PCAN, a prototype-based network that identifies ambiguous micro-action samples via a body/action hierarchy and calibrates them with contrastive prototype losses, evaluated on MA-52.
Problem
Micro-action recognition suffers from inherent ambiguity: many fine-grained categories with subtle visual differences are frequently confused, hurting accuracy.
Approach
Uses a hierarchical body/action action-tree to split misclassified samples into false-negative and false-positive ambiguous sets, then applies prototypical contrastive calibration (pull FN toward prototypes, push FP away), a prototype diversity amplification loss, and prototype-guided prediction rectification.
Key results
- PCAN achieves F1_mean 69.97% on MA-52, versus 66.66% for PoseConv3D (the strongest baseline).
- Action-level F1_macro reaches 53.83%, 6.63 points above PoseConv3D.
- Ablation on FN weights shows best results when weights decrease step by step (1, 0.5, 0.1).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply PCAN rather than PoseConv3D | PoseConv3D on MA-52 · MA-52 test | action F1_macro: 47.20 -> 53.83 | Helped |
| Apply PCAN rather than PoseConv3D | PoseConv3D on MA-52 · MA-52 test | F1_mean: 66.66 -> 69.97 | Helped |
| Apply PCAN rather than PoseConv3D | PoseConv3D on MA-52 hard-level categories · MA-52 hard-level action categories | Top-1 accuracy: +11.8% | Helped |
| Set λ=1 instead of λ=0 in the hierarchical probability loss (ablation) | PCAN RGB branch · MA-52 | action F1_macro: 52.69 -> 53.83 | Helped |
| Set λ=0.1 instead of λ=1 in the hierarchical probability loss (ablation) | PCAN RGB branch · MA-52 | action F1_macro: 53.83 -> 52.28 | Hurt |
| Set λ=10 instead of λ=1 in the hierarchical probability loss (ablation) | PCAN RGB branch · MA-52 | action F1_macro: 53.83 -> 51.82 | Hurt |
| Use α₁=1, α₂=0.5, α₃=0.1 rather than α₁=α₂=α₃=1 (ablation) | PCAN ambiguous-sample calibration · MA-52 | F1_mean: 69.67 -> 69.97 | Helped |
| Use α₁=1, α₂=0.1, α₃=0.1 rather than α₁=1, α₂=0.5, α₃=0.1 (ablation) | PCAN ambiguous-sample calibration · MA-52 | F1_mean: 69.97 -> 69.31 | Hurt |
| Use α₁=0.1, α₂=1, α₃=0.5 rather than α₁=1, α₂=0.5, α₃=0.1 (ablation) | PCAN ambiguous-sample calibration · MA-52 | F1_mean: 69.97 -> 69.21 | Hurt |
| Use α₁=1, α₂=0.5, α₃=1 rather than α₁=1, α₂=0.5, α₃=0.1 (ablation) | PCAN ambiguous-sample calibration · MA-52 | F1_mean: 69.97 -> 69.00 | Hurt |
| Use α₁=1, α₂=0.1, α₃=0.5 rather than α₁=1, α₂=0.5, α₃=0.1 (ablation) | PCAN ambiguous-sample calibration · MA-52 | F1_mean: 69.97 -> 69.50 | Hurt |
Surprising findings
- Skeleton-based GCN methods underperform 2D CNN methods on micro-action recognition, attributed to pose estimator quality and short, low-amplitude motions.
Limitations
- Ambiguous-sample categorization relies on the preliminary predictions of the backbone, so errors propagate.
- Some hard instances remain misclassified, as shown in qualitative cases with highly similar motion patterns.
Experimental setup
Datasets: MA-52
Benchmarks: MA-52
Metrics: Top-1 accuracy Top-5 accuracy F1 macro F1 micro F1_mean
Baselines: TSN TIN TSM MANet C3D I3D SlowFast VideoSwin TimeSformer UniFormer STGCN 2s-AGCN
Builds on: PoseConv3D prototypical contrastive learning exponential moving average prototype update
Closest papers by meaning
- MA-Bench: Towards Fine-grained Micro-Action Understanding (CVPR 2026 · similarity 0.88)
- SeFAR: Semi-supervised Fine-grained Action Recognition with Temporal Perturbation and Learning Stabilization (AAAI 2025 · similarity 0.84)
- Are We Making Progress in Multimodal Domain Generalization? A Comprehensive Benchmark Study (NeurIPS 2026 · similarity 0.84)
- ActionAtlas: A VideoQA Benchmark for Domain-specialized Action Recognition (NeurIPS 2024 · similarity 0.84)
- LLaVAction: evaluating and training multi-modal large language models for action understanding (ICLR 2026 · similarity 0.83)
- Storyboard-guided Alignment for Fine-grained Video Action Recognition (NeurIPS 2025 · 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.