Research map / Video / Action recognition
ALBAR: Adversarial Learning approach to mitigate Biases in 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 ALBAR, an end-to-end adversarial training method that penalizes action predictions on static (repeated-frame) clips using negated cross-entropy, entropy maximization, and a gradient penalty, to mitigate background and foreground bias in video action recognition without bias labels.
Problem
Video action recognition models rely on static background and foreground (appearance) cues rather than motion, and existing debiasing methods need bias attribute labels or critic models or focus mainly on background bias.
Approach
Feed a single frame repeated into a static clip through the same 3D encoder, apply a negative cross-entropy adversarial loss on its predictions, add an entropy-maximization loss to make static predictions uniform (preventing label-flipping collapse), and a gradient-norm penalty on the static input for stability; no bias labels or separate critic are needed.
Key results
- On HMDB51 SCUBA/SCUFO, ALBAR reaches 53.22% contrasted accuracy versus 40.28% for the best prior method (StillMix), an improvement of over 12 points.
- On the updated UCF101 protocol, ALBAR reaches 26.23% contrasted accuracy versus 24.40% for StillMix, while the combined ALBAR + StillMix augmentation reaches 31.14%.
- Using ALBAR-trained features improves downstream anomaly detection on UCF_Crime (AUC 82.39 to 84.91) and THUMOS14 localization (mAP 54.89 to 55.20).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the adversarial static-clip loss (ablation) | ALBAR on HMDB51 · SCUFO | Avg SCUFO accuracy: 20.46 -> 2.97 | Helped |
| Add the entropy maximization loss (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 42.29 -> 46.72 | Helped |
| Add the gradient penalty loss alongside the other objectives (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 46.72 -> 53.22 | Helped |
| Add the gradient penalty by itself to the baseline objective (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 27.84 -> 29.24 | No clear effect |
| Use the middle rather than a random static frame (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.02 -> 53.22 | Helped |
| Select the first rather than middle frame as the static clip (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.22 -> 50.91 | Hurt |
| Select the last rather than middle frame as the static clip (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.22 -> 49.40 | Hurt |
| Set adversarial loss weight to 0.5 rather than 1 (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.02 -> 46.56 | Hurt |
| Set adversarial loss weight to 2 rather than 1 (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.02 -> 44.45 | Hurt |
| Set adversarial loss weight to 4 rather than 1 (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.02 -> 27.76 | Hurt |
| Set entropy loss weight to 2 rather than 4 (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.02 -> 51.71 | No clear effect |
| Set entropy loss weight to 8 rather than 4 (ablation) | ALBAR on HMDB51 · contrasted accuracy | contrasted accuracy: 53.02 -> 49.26 | Hurt |
Surprising findings
- Adversarial loss alone collapses into a trivial label-flipping solution (HMDB51 contrasted accuracy 42.29 with 0.00 confl-FG), so entropy maximization is needed.
- On the original UCF101 protocol, the background leakage from bounding boxes inflates SCUBA scores, and ALBAR performs worse than some baselines there.
Limitations
- Evaluation limited to established background and foreground bias protocols; does not cover harmful demographic biases.
- Assumes no static information should be useful, which may not hold for all architectures or tasks.
- Training can be unstable and sensitive to loss weights.
Experimental setup
Datasets: Kinetics400 UCF101 HMDB51 SCUBA SCUFO ARAS Places365 THUMOS14 UCF_Crime Waterbirds
Benchmarks: HMDB51 UCF101 Kinetics400 SCUBA SCUFO ARAS UCF_Crime THUMOS14
Metrics: Top-1 accuracy contrasted accuracy frame-level AUC mAP
Baselines: Mixup VideoMix SDN BE ActorCutMix FAME StillMix
Builds on: cross-entropy adversarial debiasing WGAN-GP gradient penalty SCUBA/SCUFO protocol StillMix
Closest papers by meaning
- Rethinking Invariance Regularization in Adversarial Training to Improve Robustness-Accuracy Trade-off (ICLR 2025 · similarity 0.84)
- DVANet: Disentangling View and Action Features for Multi-View Action Recognition (AAAI 2024 · similarity 0.83)
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.83)
- VA-AR: Learning Velocity-Aware Action Representations with Mixture of Window Attention (AAAI 2025 · similarity 0.83)
- DAFA: Distance-Aware Fair Adversarial Training (ICLR 2024 · similarity 0.83)
- Bias for Action: Video Implicit Neural Representations with Bias Modulation (CVPR 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.