Research map / Video / Video segmentation and tracking
Attention Clusters: Purely Attention Based Local Feature Integration for Video Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 a shifting operation to an FC1 attention cluster (N=32) (ablation) | Flash-MNIST attention clustering · Flash-MNIST | accuracy: 83.3 -> 87.1 | Helped |
| Use attention weighting instead of global averaging (N=32) (ablation) | Flash-MNIST attention clustering · Flash-MNIST | accuracy: 2.4 -> 83.3 | Helped |
| Increase FC1 attention cluster size from N=1 to N=32 with shifting (ablation) | Flash-MNIST attention clustering · Flash-MNIST | accuracy: 51.2 -> 87.1 | Helped |
| Use FC2 rather than FC1 attention weighting at N=32 with shifting (ablation) | Flash-MNIST attention clustering · Flash-MNIST | accuracy: 87.1 -> 85.6 | Hurt |
| Add shifting to RGB attention clusters, N=8 (ablation) | Kinetics RGB classification · Kinetics validation RGB | Top-1 accuracy: 74.1 -> 74.4 | Helped |
| Add shifting to Flow attention clusters, N=64 (ablation) | Kinetics Flow classification · Kinetics validation Flow | Top-1 accuracy: 65.1 -> 67.5 | Helped |
| Add shifting to Audio attention clusters, N=128 (ablation) | Kinetics Audio classification · Kinetics validation Audio | Top-1 accuracy: 22.5 -> 24.2 | Helped |
| Increase cluster size without shifting from N=8 to N=128 (ablation) | Kinetics RGB attention clustering · Kinetics validation RGB | Top-1 accuracy: 74.1 -> 73.2 | Mixed |
| Use equal 128-unit clusters for all modalities instead of RGB 64, Flow 32, Audio 32 (ablation) | Kinetics multimodal attention clustering · Kinetics validation multimodal | Top-1 accuracy: 79.4 -> 79.3 | Hurt |
| Use FC2 instead of FC1 at large cluster sizes (ablation) | Flash-MNIST attention clustering · Flash-MNIST | accuracy: 83.3 -> 83.2 | Hurt |
| Use multimodal attention clusters rather than Bi-directional LSTM fusion | Kinetics RGB+Flow+Audio fusion · Kinetics validation | Top-1 accuracy: 78.2 -> 79.4 | Helped |
| Use attention clusters rather than TSN for RGB classification | Kinetics RGB classification · Kinetics validation RGB | Top-1 accuracy: 73.0 -> 75.0 | Helped |
Closest papers by meaning
- Multimodal Keyless Attention Fusion for Video Classification (AAAI 2018 · similarity 0.88)
- Recurrent Mixture Density Network for Spatiotemporal Visual Attention (ICLR 2017 · similarity 0.87)
- Is Space-Time Attention All You Need for Video Understanding? (ICML 2021 · similarity 0.87)
- MCA: Moment Channel Attention Networks (AAAI 2024 · similarity 0.85)
- HyperVAttention: Efficient Sparse Attention with Spatio-Temporal Clustering for Video Diffusion (NeurIPS 2026 · similarity 0.85)
- Degrees of Freedom for Linear Attention: Distilling Softmax Attention with Optimal Feature Efficiency (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.