Research map / Video / Action recognition
Amplify, Aggregate, and Adjust: VideoMAE-based Holistic-Subtle Aggregation for Micro-Action Recognition
Yan Zhang, Nan Pu, Wenjing Li, Zhun Zhong, Meng Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TGLA subtle-motion branch to VideoMAE baseline (ablation) | VideoMAE baseline · MA-52 | Action Top-1: 66.47 -> 67.04 | Improved |
| add HSMA dual cross-attention on top of TGLA (ablation) | VideoMAE + TGLA · MA-52 | Action Top-1: 67.04 -> 67.63 | Improved |
| add CADA on top of TGLA + HSMA (ablation) | VideoMAE + TGLA + HSMA · MA-52 | Action Top-1: 67.63 -> 67.92 | Improved |
| add TGLA to baseline (macro F1) (ablation) | VideoMAE baseline · MA-52 | Action F1 macro: 52.46 -> 55.09 | Improved |
| use only selective confidence distillation L_SCD in CADA (ablation) | A3-MAE without CADA · MA-52 | Action Top-1: 67.63 -> 67.49 | Hurt |
| add confidence-weight alignment L_CWA to L_SCD (ablation) | CADA · MA-52 | Action Top-1: 67.49 -> 67.60 | Improved |
| add weight entropy regularization L_WER to L_SCD (ablation) | CADA · MA-52 | Action Top-1: 67.49 -> 67.69 | Improved |
| use all three CADA losses instead of L_SCD + L_WER (ablation) | CADA · MA-52 | Action Top-1: 67.69 -> 67.92 | Improved |
| use aggregated head Ga instead of single-branch head (ablation) | A3-MAE · MA-52 | Action Top-1: 67.60 -> 67.92 | Improved |
| use instance-adaptive CADA weights instead of fixed subtle-branch weight (ablation) | A3-MAE · MA-52 | Action Top-1: every fixed setting underperforms CADA (67.92) | Improved |
| use 16 frames instead of fewer frames per video (ablation) | A3-MAE · MA-52 | Action Top-1: best results achieved at N = 16 | Improved |
| raise TG-mask threshold from 90 to larger values (ablation) | A3-MAE · MA-52 | Action Top-1: 67.92 -> 67.08 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- VideoMAE — VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- MMAD: Multi-label Micro-Action Detection in Videos (ICCV 2025 · similarity 0.91)
- Prototypical Calibrating Ambiguous Samples for Micro-Action Recognition (AAAI 2025 · similarity 0.89)
- UCF-STAR: A Large Scale Still Image Dataset for Understanding Human Actions (AAAI 2020 · similarity 0.88)
- AVA: A Video Dataset of Spatio-Temporally Localized Atomic Visual Actions (CVPR 2018 · similarity 0.87)
- MA-Bench: Towards Fine-grained Micro-Action Understanding (CVPR 2026 · similarity 0.86)
- Chirality in Action: Time-Aware Video Representation Learning by Latent Straightening (NeurIPS 2025 · similarity 0.85)
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.
