Research map / Video / Action recognition
Is Appearance Free Action Recognition Possible?
Filip Ilic, Thomas Pock, Richard P. Wildes
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 |
|---|---|---|---|
| remove appearance (use AFD instead of RGB) | 11 action architectures · UCF101 vs AFD101 | top-1 accuracy: approximately 30% average drop | Hurt |
| add explicit optical flow stream to X3D M (E2S-X3D) | X3D M · AFD101 | top-1 accuracy: 58.2 -> 73.9 | Improved |
| add explicit optical flow stream to X3D M (E2S-X3D) | X3D M · UCF101 | top-1 accuracy: 81.5 -> 85.7 | Improved |
| add RAFT flow stream to X3D M | X3D M · AFD5 | top-1 accuracy: 71.5 -> 90.8 | Improved |
| feed RGB to both streams instead of RGB and optical flow (ablation) | E2S-X3D · UCF101 | top-1 accuracy: 10.6% drop when removed | Improved |
| feed RGB to both streams instead of RGB and optical flow (ablation) | E2S-X3D · AFD101 | top-1 accuracy: 30.7% drop when removed | Improved |
| use X3D S for flow stream instead of X3D M (ablation) | E2S-X3D · AFD5 | top-1 accuracy: 78.9 -> 80.3 (flow-only) | Improved |
| use X3D M for RGB stream instead of S (ablation) | E2S-X3D · AFD5 | top-1 accuracy: 69.8 -> 71.5 (RGB-only) | Improved |
| prioritize temporal information (Fast stream, I3D OF) | action architectures · UCF101 vs AFD101 | top-1 accuracy drop: Fast 4.6 vs I3D 24.5 drop | Improved |
| Kinetics pretraining before AFD finetuning | architectures · AFD | top-1 accuracy | No clear effect |
| nearest-neighbor instead of bilinear interpolation in warping (ablation) | AFD generation · AFD | leakage of single-frame information | Improved |
Closest papers by meaning
- Recognize Actions by Disentangling Components of Dynamics (CVPR 2018 · similarity 0.87)
- Scene Flow to Action Map: A New Representation for RGB-D Based Action Recognition With Convolutional Neural Networks (CVPR 2017 · similarity 0.86)
- Quantifying and Learning Static vs. Dynamic Information in Deep Spatiotemporal Networks (TPAMI 2024 · similarity 0.86)
- TrAction: Action Recognition with Sparse Trajectories (NeurIPS 2026 · similarity 0.85)
- VA-AR: Learning Velocity-Aware Action Representations with Mixture of Window Attention (AAAI 2025 · similarity 0.84)
- Disentangled Concepts Speak Louder Than Words: Explainable Video Action Recognition (NeurIPS 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.
