Research map / Video / Action recognition
Gen2Balance: Generative Balancing for Long-Tailed Video Action Recognition
Prajwal Gatti, Simon Jenni, Fabian Caba, Dima Damen
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 generated videos with plain CE (ablation) | CE on K100-LT · K100-LT (B=330) | avg class accuracy: 54.8 -> 50.8 | Hurt |
| use Balanced Softmax loss with generated data (ablation) | CE on augmented data · K100-LT (B=330) | avg class accuracy: 50.8 -> 61.3 | Improved |
| add stage 2 fine-tuning on real data (ablation) | BSCE stage 1 (N+N') · K100-LT (B=330) | avg class accuracy: 61.3 -> 70.6 | Improved |
| use real class frequencies for BSCE margins instead of augmented frequencies (ablation) | two-stage Gen2Balance · K100-LT (B=330) | few-shot accuracy: 61.0 -> 62.1 | Improved |
| train stage 1 longer (135 epochs) instead of two stages (ablation) | BSCE real freq · K100-LT (B=330) | avg class accuracy: 67.8 -> 67.9 | No clear effect |
| add diversified prompts over naive prompting (ablation) | WAN generation · K100-LT 10 ablated classes | accuracy on 10 classes: 37.3 -> 44.9 | Improved |
| add action profile (ablation) | WAN generation · K100-LT 10 ablated classes | accuracy on 10 classes: 44.9 -> 47.2 | Improved |
| add in-context exemplars to action profile (ablation) | WAN generation · K100-LT 10 ablated classes | accuracy on 10 classes: 47.2 -> 49.8 | Improved |
| use generated videos instead of web-retrieved real videos (ablation) | augmentation source · K100-LT (B=330) | avg class accuracy: 66.9 -> 70.9 | Improved |
| use generated videos instead of generated images (ablation) | augmentation source · K100-LT (B=330) | avg class accuracy: 68.0 -> 70.9 | Improved |
| use naive templated prompts (ablation) | WAN generation · K100-LT (B=330) | avg class accuracy: 70.9 -> 65.1 | Hurt |
| raise filling threshold from B=330 to B=990 (ablation) | Gen2Balance · K100-LT | avg class accuracy: 70.9 -> 72.6 | Improved |
Closest papers by meaning
- TimeBalance: Temporally-Invariant and Temporally-Distinctive Video Representations for Semi-Supervised Action Recognition (CVPR 2023 · similarity 0.86)
- Scaling Autoregressive Video Models (ICLR 2020 · similarity 0.86)
- Align Before Adapt: Leveraging Entity-to-Region Alignments for Generalizable Video Action Recognition (CVPR 2024 · similarity 0.85)
- ActionAtlas: A VideoQA Benchmark for Domain-specialized Action Recognition (NeurIPS 2024 · similarity 0.85)
- On Discriminative vs. Generative classifiers: Rethinking MLLMs for Action Understanding (ICLR 2026 · similarity 0.84)
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · similarity 0.84)
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.
