Research map / Video / Video segmentation and tracking
Rethinking Video ViTs: Sparse Video Tubes for Joint Image and Video Learning
Read the paperCheck 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 |
|---|---|---|---|
| Jointly train on ImageNet and Kinetics rather than Kinetics alone | TubeViT-L · Kinetics-600 | Top-1 accuracy: 85.6 -> 91.5 | Helped |
| Jointly train on both datasets rather than train sequentially | TubeViT-L · Kinetics-600 | Top-1 accuracy: 90.4 -> 91.5 | Helped |
| Use sparse video tubes rather than 2D patches only | TubeViT-L · Kinetics-600 | Top-1 accuracy: 87.6 -> 91.5 | Helped |
| Add video tubes rather than use inflated 3D patches | TubeViT-L · Kinetics-600 | Top-1 accuracy: 88.4 -> 91.5 | Helped |
| Jointly train with Kinetics videos | ViT-B · ImageNet-1k | accuracy: 78.1 -> 81.4 | Helped |
| Add the gated raw-tube connection (ablation) | Image-to-video scaling with ViT-H · Kinetics-600 | accuracy: 86.8 -> 89.7 | Helped |
| Use fixed cosine positions that encode tube strides instead of no positional embeddings (ablation) | TubeViT-B · Kinetics-600 | Top-1 accuracy: 78.6 -> 84.5 | Helped |
| Encode stride in fixed cosine positional embeddings (ablation) | TubeViT-B · Kinetics-600 | Top-1 accuracy: 77.7 -> 84.5 | Helped |
| Increase the number of tubes from one to eight (ablation) | TubeViT-B · Kinetics-600 | Top-1 accuracy: 78.4 -> 85.4 | Helped |
| Add space-to-depth x4 spatially (ablation) | TubeViT-B · Kinetics-600 | Top-1 accuracy: 83.4 -> 85.4 | Helped |
| Increase evaluation tokens from 559 to 768 (ablation) | TubeViT-B · Kinetics-600 | Top-1 accuracy: 84.5 -> 84.9 | Helped |
| Increase evaluation tokens from 559 to 1536 (ablation) | TubeViT-B · Kinetics-600 | Top-1 accuracy: 84.5 -> 83.5 | Hurt |
Closest papers by meaning
- Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words (ICLR 2024 · similarity 0.86)
- TTVidT: Decoupling the Temporal Axis for Efficient Motion-Centric Video Pretraining (NeurIPS 2026 · similarity 0.85)
- Memory Consolidation Enables Long-Context Video Understanding (ICML 2024 · similarity 0.85)
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.85)
- VideoPoet: A Large Language Model for Zero-Shot Video Generation (ICML 2024 · similarity 0.84)
- VONet: Unsupervised Video Object Learning With Parallel U-Net Attention and Object-wise Sequential VAE (ICLR 2024 · 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.