Research map / Video / Action recognition
Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset
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 inflated 3D convolutions in a two-stream RGB-and-flow model | Inception-v1 action recognition · UCF-101 split 1 | accuracy (RGB + Flow): 91.2 -> 93.4 | Helped |
| Use inflated 3D convolutions in a two-stream RGB-and-flow model | Inception-v1 action recognition · HMDB-51 split 1 | accuracy (RGB + Flow): 58.3 -> 66.4 | Helped |
| Use inflated 3D convolutions in a two-stream RGB-and-flow model | Inception-v1 action recognition · Kinetics | Top-1 accuracy (RGB + Flow): 67.2 -> 74.2 | Helped |
| Pre-train I3D on Kinetics before fine-tuning | Two-Stream I3D · UCF-101 split 1 | accuracy: 93.4 -> 98.0 | Helped |
| Pre-train I3D on Kinetics before fine-tuning | Two-Stream I3D · HMDB-51 split 1 | accuracy: 66.4 -> 81.2 | Helped |
| Add an optical-flow stream to RGB I3D (ablation) | Two-Stream I3D · Kinetics | Top-1 accuracy: 71.1 -> 74.2 | Helped |
| Add an optical-flow stream to RGB I3D (ablation) | Two-Stream I3D · UCF-101 split 1 | accuracy: 84.5 -> 93.4 | Helped |
| Add an optical-flow stream to RGB I3D (ablation) | Two-Stream I3D · HMDB-51 split 1 | accuracy: 49.8 -> 66.4 | Helped |
| Initialize I3D with ImageNet-pretrained weights before Kinetics training (ablation) | Two-Stream I3D RGB stream · Kinetics | Top-1 accuracy: 68.4 -> 71.1 | Helped |
| Initialize I3D with ImageNet-pretrained weights before Kinetics training (ablation) | Two-Stream I3D Flow stream · Kinetics | Top-1 accuracy: 61.5 -> 63.4 | Helped |
| Initialize I3D with ImageNet-pretrained weights before Kinetics training (ablation) | Two-Stream I3D · Kinetics | Top-1 accuracy (RGB + Flow): 71.6 -> 74.2 | Helped |
| Fine-tune the full Kinetics-pretrained I3D instead of training only its final layer (ablation) | Two-Stream I3D · UCF-101 split 1 | accuracy: 97.7 -> 98.0 | Helped |
Closest papers by meaning
- Can Spatiotemporal 3D CNNs Retrace the History of 2D CNNs and ImageNet? (CVPR 2018 · similarity 0.89)
- T-C3D: Temporal Convolutional 3D Network for Real-Time Action Recognition (AAAI 2018 · similarity 0.88)
- A Structured Model for Action Detection (CVPR 2019 · similarity 0.87)
- Deep Analysis of CNN-Based Spatio-Temporal Representations for Action Recognition (CVPR 2021 · similarity 0.87)
- VideoNet: A Large-Scale Dataset for Domain-Specific Action Recognition (CVPR 2026 · similarity 0.87)
- More Is Less: Learning Efficient Video Representations by Big-Little Network and Depthwise Temporal Aggregation (NeurIPS 2019 · 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.