Research map / Video / Action recognition
TAda! Temporally-Adaptive Convolutions for Video Understanding
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 TAdaConv to existing video backbones | SlowOnly, SlowFast, R(2+1)D and R3D · Kinetics-400 and Something-Something-V2 | Top-1: +1.29 K400; +2.99 SSV2 (SlowOnly 8×8) | Helped |
| Use temporally varying calibration weights instead of shared calibration weights (ablation) | TAdaConv · SSV2 | Top-1: 34.3 -> 45.4 | Helped |
| Generate dynamic calibration weights rather than learnable weights (ablation) | TAdaConv · SSV2 | Top-1: 45.4 -> 53.8 | Helped |
| Generate calibration from local and global temporal context (ablation) | TAdaConv · SSV2 | Top-1: 53.8 -> 59.2 | Helped |
| Initialize calibration generation so the layer starts as the pretrained convolution (ablation) | TAdaConv · SSV2 | Top-1: 53.8 -> 47.8 (without proposed initialization) | Helped |
| Add local temporal context to weight generation using a larger kernel (ablation) | TAdaConv · SSV2 | Top-1: 37.5 -> 56.5 | Helped |
| Use a two-layer non-linear calibration generator instead of a linear generator (ablation) | TAdaConv · SSV2 | Top-1: 57.3 -> 57.8 | Helped |
| Add a shortcut branch to average feature aggregation (ablation) | TAda2D · SSV2 | Top-1: 60.1 -> 61.5 | Helped |
| Use separate batch normalization for shortcut and aggregation branches (ablation) | TAda2D · SSV2 | Top-1: 61.5 -> 63.8 | Helped |
| Replace average pooling with max pooling (ablation) | TAda2D feature aggregation · SSV2 | Top-1: 63.8 -> 63.5 | Hurt |
| Replace average pooling with mixed average and max pooling (ablation) | TAda2D feature aggregation · SSV2 | Top-1: 63.8 -> 63.7 | Hurt |
| Calibrate output channels rather than input channels (ablation) | TAdaConv · SSV2 | Top-1: 63.8 -> 63.4 | Hurt |
Closest papers by meaning
- CDC: Convolutional-De-Convolutional Networks for Precise Temporal Action Localization in Untrimmed Videos (CVPR 2017 · similarity 0.88)
- Temporal Convolutional Networks for Action Segmentation and Detection (CVPR 2017 · similarity 0.88)
- Temporal Residual Networks for Dynamic Scene Recognition (CVPR 2017 · similarity 0.87)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.87)
- End-to-End Temporal Action Detection with 1B Parameters Across 1000 Frames (CVPR 2024 · similarity 0.86)
- DejaVid: Encoder-Agnostic Learned Temporal Matching for Video Classification (CVPR 2025 · similarity 0.86)
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.