Research map / Video / Action recognition
2D or not 2D? Adaptive 3D Convolution Selection for Efficient Video Recognition
Check 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 |
|---|---|---|---|
| Use adaptive frame and 3D-convolution policies instead of using every frame and convolution | I3D ResNet-50 · FCVID, 8-frame per clip | GFLOPs: 58.6 -> 35.6 | Helped |
| Use adaptive frame and 3D-convolution policies instead of using every frame and convolution | I3D ResNet-50 · ActivityNet, 8-frame per clip | mAP: 82.6 -> 82.6 | No clear effect |
| Use adaptive frame and 3D-convolution policies instead of using every frame and convolution | I3D ResNet-50 · Mini-Kinetics, 8-frame per clip | accuracy: 79.0 -> 78.9 | No clear effect |
| Learn policies conditioned on each input rather than sample random policies at similar compute | Ada3D · FCVID, 8-frame per clip | mAP: 78.1 -> 81.9 | Helped |
| Learn policies conditioned on each input rather than use random policies with joint fine-tuning | Ada3D · ActivityNet, 8-frame per clip | mAP: 81.1 -> 82.6 | Helped |
| Transfer learned policies from Mini-Kinetics to Kinetics | I3D ResNet-50 · Kinetics validation | accuracy: 73.1 -> 72.8 | No clear effect |
| Add AdaFrame clip selection to Ada3D | Ada3D · FCVID, 10 clips | mAP: 81.9 -> 82.0 | No clear effect |
| Add initial selection-network training before joint fine-tuning (ablation) | Ada3D · FCVID, 3-clip test | mAP: 76.5 -> 77.8 | Helped |
| Add joint fine-tuning of the video model and selection network (ablation) | Ada3D · FCVID, 3-clip test | mAP: 75.9 -> 77.8 | Helped |
| Add initial selection-network training before joint fine-tuning (ablation) | Ada3D · ActivityNet, 3-clip test | mAP: 75.1 -> 76.1 | Helped |
| Add joint fine-tuning of the video model and selection network (ablation) | Ada3D · ActivityNet, 3-clip test | mAP: 74.3 -> 76.1 | Helped |
| Learn both frame and 3D-convolution policies instead of using random policies for both (ablation) | Ada3D · FCVID, 3-clip test | mAP: 75.5 -> 77.8 | Helped |
Closest papers by meaning
- Optimization Planning for 3D ConvNets (ICML 2021 · similarity 0.88)
- Omni3D: A Large Benchmark and Model for 3D Object Detection in the Wild (CVPR 2023 · similarity 0.87)
- T-C3D: Temporal Convolutional 3D Network for Real-Time Action Recognition (AAAI 2018 · similarity 0.87)
- X3D: Expanding Architectures for Efficient Video Recognition (CVPR 2020 · similarity 0.87)
- Learning from Unlabelled Videos Using Contrastive Predictive Neural 3D Mapping (ICLR 2020 · similarity 0.87)
- SP3D: Boosting Sparsely-Supervised 3D Object Detection via Accurate Cross-Modal Semantic Prompts (CVPR 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.