Research map / Video / Video-language understanding
TubeDETR: Spatio-Temporal Video Grounding With Transformers
Check 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 |
|---|---|---|---|
| Add time encoding and temporal self-attention to the space-only decoder (ablation) | TubeDETR · VidSTG validation | vIoU@0.3: +17.9% on $vIoU@0.3$ | Helped |
| Add time encoding to the space-only decoder (ablation) | TubeDETR · VidSTG validation | vIoU@0.3: +13.4% on $vIoU@0.3$ | Helped |
| Add temporal self-attention to a decoder with time encoding (ablation) | TubeDETR · VidSTG validation | vIoU@0.3: +4.5% on $vIoU@0.3$ | Helped |
| Initialize with MDETR pretrained weights instead of ImageNet visual-backbone initialization (ablation) | TubeDETR · VidSTG validation | vIoU@0.3: +8.1% on $vIoU@0.3$ | Helped |
| Transfer MDETR decoder spatial self-attention weights to temporal self-attention (ablation) | TubeDETR · VidSTG validation | vIoU@0.3: +3.3% on $vIoU@0.3$ | Helped |
| Increase frame resolution from 224 to 288 at stride 2 (ablation) | TubeDETR · VidSTG validation | GPU memory: 16.2GB vs 23.7GB | Hurt |
| Use temporal stride 4 and resolution 352 instead of stride 1 and resolution 224 (ablation) | TubeDETR video-text encoder · VidSTG and HC-STVG2.0 validation | spatio-temporal grounding performance and GPU memory: 16.2GB vs 23.7GB | Helped |
| Add the fast visual-only branch to the slow branch (ablation) | TubeDETR video-text encoder · VidSTG validation | vIoU@0.3: +2.1% $vIoU@0.3$ | Helped |
| Add the fast visual-only branch to the slow branch (ablation) | TubeDETR video-text encoder · HC-STVG2.0 validation | vIoU@0.3: +2.9% $vIoU@0.3$ | Helped |
| Add the slow multi-modal branch to the fast-only encoder (ablation) | TubeDETR video-text encoder · VidSTG validation | vIoU@0.3 | Helped |
| Preserve spatially detailed features in the fast branch instead of spatially pooling them (ablation) | TubeDETR video-text encoder · VidSTG validation | spatio-temporal grounding performance | Helped |
| Replace sum-plus-linear aggregation with sigmoid gating (ablation) | TubeDETR video-text encoder · VidSTG validation | spatio-temporal grounding performance | Hurt |
Closest papers by meaning
- Knowing Your Target: Target-Aware Transformer Makes Better Spatio-Temporal Video Grounding (ICLR 2025 · similarity 0.89)
- End-to-End Spatio-Temporal Action Localisation with Video Transformers (CVPR 2024 · similarity 0.88)
- STPro: Spatial and Temporal Progressive Learning for Weakly Supervised Spatio-Temporal Grounding (CVPR 2025 · similarity 0.87)
- CoSTA: End-to-End Comprehensive Space-Time Entanglement for Spatio-Temporal Video Grounding (AAAI 2024 · similarity 0.87)
- Unleashing the Potential of Multimodal LLMs for Zero-Shot Spatio-Temporal Video Grounding (NeurIPS 2025 · similarity 0.87)
- Contextual Self-paced Learning for Weakly Supervised Spatio-Temporal Video Grounding (ICLR 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.