Research map / 3D vision / Human pose estimation
Bringing Image Scene Structure to Video via Frame-Clip Consistency of Object Tokens
Elad Ben Avraham, Roei Herzig, Karttikeya Mangalam, Amir Bar, Anna Rohrbach, Leonid Karlinsky, Trevor Darrell, Amir Globerson
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 |
|---|---|---|---|
| Use SViT with annotated images from a different domain | MViTv2 on SomethingElse · SomethingElse | Top-1 accuracy: 63.3 -> 65.1 | Improved |
| Use only 2% of the in-domain image annotations (ablation) | SViT on SomethingElse · SomethingElse | Top-1 accuracy: 63.3 -> 65.6 | Improved |
| Finetune jointly on video and auxiliary images rather than finetuning only on video | SViT on SomethingElse · SomethingElse | Top-1 accuracy: 64.2 -> 65.1 | Improved |
| Add object tokens and the SViT loss to the multi-task baseline (ablation) | MViTv2 on SomethingElse · SomethingElse | Top-1 accuracy: 63.8 -> 65.0 | Improved |
| Add frame-clip consistency loss alongside object tokens (ablation) | MViTv2 on SomethingElse · SomethingElse | Top-1 accuracy: 65.0 -> 65.8 | Improved |
| Increase annotated image use from 2% to 25% (ablation) | SViT-ID on SomethingElse · SomethingElse | Top-1 accuracy: 65.6 -> 65.6 | No clear effect |
| Increase annotated image use from 2% to 50% (ablation) | SViT-ID on SomethingElse · SomethingElse | Top-1 accuracy: 65.6 -> 65.7 | No clear effect |
| Use real HAOG annotations instead of random HAOGs (ablation) | SViT-ID on SomethingElse · SomethingElse | Top-1 accuracy: 50.6 -> 65.8 | Improved |
| Use SViT rather than MViTv2 on SSv2 | MViTv2 · SSv2 | Top-1 accuracy: 68.2 -> 69.7 | Improved |
| Use SViT rather than MViTv2 on Diving48 | MViTv2 · Diving48 | Top-1 accuracy: 73.1 -> 79.8 | Improved |
| Use SViT rather than MViTv2 on AVA-V2.2 | MViTv2 · AVA-V2.2 | mAP: 26.8 -> 28.5 | Improved |
| Apply SViT on clips with no annotated hands (ablation) | MViTv2 on SomethingElse · SomethingElse, no annotated hands | Top-1 accuracy: 64.6 -> 66.5 | Improved |
Closest papers by meaning
- Optimizing Factorized Encoder Models: Time and Memory Reduction for Scalable and Efficient Action Recognition (ECCV 2024 · similarity 0.88)
- Object-Region Video Transformers (CVPR 2022 · similarity 0.88)
- How Can Objects Help Action Recognition? (CVPR 2023 · similarity 0.88)
- Graph Convolutional Module for Temporal Action Localization in Videos (TPAMI 2021 · similarity 0.88)
- VONet: Unsupervised Video Object Learning With Parallel U-Net Attention and Object-wise Sequential VAE (ICLR 2024 · similarity 0.86)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · 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.
