Research map / Video / Video segmentation and tracking
Unsupervised Learning of Long-Term Motion Dynamics for Videos
Zelun Luo, Boya Peng, De-An Huang, Alexandre Alahi, Li Fei-Fei
Check 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 |
|---|---|---|---|
| Increase motion prediction length from 3 steps to 8 steps (ablation) | Our method · NTU-RGB+D | mean AP (%): 62.5 -> 66.2 | Improved |
| Predict 3D motion rather than quantized 2D optical flow (ablation) | Our method · NTU-RGB+D | mean AP (%): 58.8 -> 66.2 | Improved |
| Pretrain the encoder to predict motion (ablation) | Our encoder architecture · NTU-RGB+D | mean AP (%): 37.5 -> 66.2 | Improved |
| Pretrain the encoder to predict motion (ablation) | Our encoder architecture · NTU-RGB+D RGB | mean AP (%): 34.1 -> 56 | Improved |
| Fine-tune the pretrained encoder for activity recognition (ablation) | Our method · NTU-RGB+D | mean AP (%) | No clear effect |
| Use a learnable codebook rather than a uniform codebook (ablation) | Atomic 3D-flow quantization · NTU-RGB+D motion prediction | activity recognition performance | Improved |
| Use a k-means codebook rather than a uniform codebook (ablation) | Atomic 3D-flow quantization · NTU-RGB+D motion prediction | activity recognition performance | Hurt |
| Use soft labels with more than one nearest codeword (ablation) | Atomic 3D-flow quantization · NTU-RGB+D motion prediction | activity recognition performance | Improved |
| Add ImageNet semantic features to the learned representation (ablation) | Our method · UCF-101 | mean AP (%): 53.0 -> 79.3 | Improved |
Closest papers by meaning
- Learning Long-term Motion Embeddings for Efficient Kinematics Generation (CVPR 2026 · similarity 0.90)
- Lattice Long Short-Term Memory for Human Action Recognition (ICCV 2017 · similarity 0.90)
- Learning Video Representations by Transforming Time (ECCV 2020 · similarity 0.89)
- Enhancing Unsupervised Video Representation Learning by Decoupling the Scene and the Motion (AAAI 2021 · similarity 0.88)
- Learning Temporal Dynamics From Cycles in Narrated Video (ICCV 2021 · similarity 0.88)
- Unsupervised Learning of View-invariant Action Representations (NeurIPS 2018 · similarity 0.88)
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.
