Research map / Video / Video segmentation and tracking
Adversarial Framework for Unsupervised Learning of Motion Dynamics in Videos
Concetto Spampinato, Simone Palazzo, Pierluca D’Oro, Daniela Giordano, Mubarak A. Shah
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 |
|---|---|---|---|
| add SLERP trajectory latent space (ablation) | VOS-GAN baseline · Golf course (video generation) | FID: 44.02 -> 41.76 | Improved |
| add LSTM encoding of the latent trajectory (ablation) | VOS-GAN Model 1 · Golf course (video generation) | FID: 41.76 -> 40.54 | Improved |
| add segmentation self-supervision loss (ablation) | VOS-GAN Model 2 · Golf course (video generation) | FID: 40.54 -> 38.79 | Improved |
| add optical flow loss (ablation) | VOS-GAN Model 3 · Golf course (video generation) | FID: 38.79 -> 33.18 | Improved |
| add alpha schedule on dense-prediction loss weight (ablation) | VOS-GAN Model 4 · Golf course (video generation) | FID: 33.18 -> 31.32 | Improved |
| add all components (SLERP, LSTM, segmentation, OF, alpha) (ablation) | VOS-GAN baseline · Golf course (video generation) | Inception score: 1.98 -> 3.16 | Improved |
| adversarial training of segmentation net instead of training on generator masks directly | VOS-GAN segmentation · DAVIS 2017 (unsupervised) | M_J: 21.87 -> 27.01 | Improved |
| adversarial pretraining before fine-tuning instead of synthetic pretraining | VOS-GAN segmentation · DAVIS 2017 (fine-tuned) | M_J: 55.12 -> 61.65 | Improved |
| adversarial pretraining then fine-tuning vs supervised-only training | VOS-GAN segmentation · DAVIS 2017 (fine-tuned) | M_J: 53.41 -> 61.65 | Improved |
| add segmentation mask self-supervision beyond optical flow only (ablation) | VOS-GAN discriminator · DAVIS 2017 (fine-tuned) | M_J: 54.32 -> 61.65 | Improved |
| use VOS-GAN video generation instead of MoCoGAN | video GAN · Weizmann Action (video generation) | Inception score: 3.76 -> 3.29 | Hurt |
| use VOS-GAN discriminator instead of MoCoGAN for transfer | video GAN · UCF101 (Weizmann-trained, transfer) | accuracy: 18.47 -> 29.42 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- MoCoGAN — MoCoGAN: Decomposing Motion and Content for Video Generation (CVPR 2018) · better on 6 of 8 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Generating Videos with Scene Dynamics (NeurIPS 2016 · similarity 0.91)
- Self-Supervised Video GANs: Learning for Appearance Consistency and Motion Coherency (CVPR 2021 · similarity 0.89)
- Unsupervised Learning of View-invariant Action Representations (NeurIPS 2018 · similarity 0.89)
- Learning Features by Watching Objects Move (CVPR 2017 · similarity 0.89)
- Hierarchical Video Generation From Orthogonal Information: Optical Flow and Texture (AAAI 2018 · similarity 0.88)
- Self-Supervised Video Representation Learning by Context and Motion Decoupling (CVPR 2021 · 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.
