Research map / Video / Video segmentation and tracking
Leaping Into Memories: Space-Time Deep Feature Synthesis
Alexandros Stergiou, Nikos Deligiannis
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 feature diversity and temporal coherence regularizers to model priming (ablation) | LEAPS · Kinetics-400 val synthesized videos (3D R50) | top-1 accuracy of inverted model (%): 67.9 -> 86.7 | Improved |
| add feature diversity regularizer to priming (ablation) | LEAPS · Kinetics-400 val synthesized videos (X3D-M) | top-1 accuracy of inverted model (%): 73.4 -> 81.1 | Improved |
| add temporal coherence regularizer to priming (ablation) | LEAPS · Kinetics-400 val synthesized videos (X3D-M) | top-1 accuracy of inverted model (%): 73.4 -> 78.0 | Improved |
| use priming from a stimulus video instead of plain DeepDream/AM | video model inversion · Kinetics-400 val synthesized videos (3D R50) | top-1 accuracy of inverted model (%): 41.4 -> 67.9 | Improved |
| use JVS instead of cosine as priming distance (ablation) | LEAPS · Kinetics-400 val synthesized videos (3D R50) | top-1 accuracy of inverted model (%): 72.1 -> 86.7 | Improved |
| use JVS instead of l2 priming distance (ablation) | LEAPS · Kinetics-400 val synthesized videos (3D R50) | top-1 accuracy of inverted model (%): 84.3 -> 86.7 | Improved |
| use priming loss on all layers instead of first 80% (ablation) | LEAPS · Kinetics-400 val synthesized videos (3D R50) | top-1 accuracy of inverted model (%): 69.3 -> 86.7 | Improved |
| increase synthesized video size from 8x224^2 to 16x224^2 (ablation) | LEAPS · Kinetics-400 val synthesized videos (3D R50) | top-1 accuracy of inverted model (%): 86.7 -> 87.0 | No clear effect |
| reduce spatial resolution of synthesized video to 8x182^2 (ablation) | LEAPS · Kinetics-400 val synthesized videos (3D R50) | top-1 accuracy of inverted model (%): 86.7 -> 84.3 | Hurt |
| apply temporal coherence regularization on all layers instead of final layer only (ablation) | LEAPS · qualitative zumba example | visual quality | Hurt |
| use multiple stimuli instead of regularizers | LEAPS · Kinetics-400 (3D R50, X3D-M) | top-1 accuracy | Mixed |
Closest papers by meaning
- Video-to-Video Synthesis (NeurIPS 2018 · similarity 0.88)
- State Space Prompting via Gathering and Spreading Spatio-Temporal Information for Video Understanding (NeurIPS 2025 · similarity 0.86)
- Learning Conditional Space-Time Prompt Distributions for Video Class-Incremental Learning (CVPR 2025 · similarity 0.86)
- VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis (ICLR 2025 · similarity 0.86)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.86)
- SpikeVideoFormer: An Efficient Spike-Driven Video Transformer with Hamming Attention and $\mathcal{O}(T)$ Complexity (ICML 2025 · 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.
