Research map / Video / Video generation
Order-Prompted Tag Sequence Generation for Video Tagging
Zongyang Ma, Ziqi Zhang, Yuxin Chen, Zhongang Qi, Yingmin Luo, Zekun Li, Chunfeng Yuan, Bing Li, Xiaohu Qie, Ying Shan, Weiming Hu
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 sample-dependent order prompts instead of frequency-ordered tag sequences (ablation) | OP-TSG · CREATE-210K | label-based F1 (all / rare / common): 28.3 -> 32.2 (rare 23.2 -> 27.8, common 43.2 -> 45.3) | Improved |
| add random shuffling of order prompts during training (ablation) | OP-TSG · CREATE-210K | label-based F1: 30.1 -> 32.2 (rare 24.8 -> 27.8; common 45.4 vs 45.3) | Improved |
| add prompt-tag contrastive learning (ablation) | OP-TSG · CREATE-210K | label-based F1: 28.4 -> 32.2 (rare 22.9 -> 27.8) | Improved |
| freeze tag embeddings rather than learning them (ablation) | OP-TSG · CREATE-210K | label-based F1: 31.4 -> 32.2 | Improved |
| use [PAD] tags rather than a binary prompt-label head (ablation) | OP-TSG · CREATE-210K | label-based F1: 29.7 -> 32.2 | Improved |
| generate tags with decoder rather than classify from each order prompt (ablation) | OP-TSG · CREATE-210K | label-based F1: 28.7 -> 32.2 | Improved |
| use both video and title input rather than video only (ablation) | OP-TSG · CREATE-210K | label-based F1: 21.8 -> 32.2 (video only), 23.3 -> 32.2 (title only) | Improved |
| number of order prompts 30 versus fewer or more (ablation) | OP-TSG · CREATE-210K | label-based F1: 29.8 (5), 32.2 (30), 31.6 (50) | Mixed |
| lower [PAD] loss weight to 0.3 (ablation) | OP-TSG · CREATE-210K | label-based F1: 31.2 (1.0) -> 32.2 (0.3); 31.4 at 0.1 | Mixed |
| deepen prompt encoder beyond 6 layers (ablation) | OP-TSG · CREATE-210K | label-based F1: 32.2 (6) vs 32.1 (8) | No clear effect |
| OP-TSG versus best baselines on common tags | OP-TSG · CREATE-210K | common-tag label F1: 45.3 vs Asy 45.8 | No clear effect |
| add CREATE-3M pretraining | OP-TSG · CREATE-210K | label-based F1 all tags: 32.2 -> 35.5 | Improved |
Closest papers by meaning
- Chinese Title Generation for Short Videos: Dataset, Metric and Algorithm (TPAMI 2024 · similarity 0.86)
- T2V-CompBench: A Comprehensive Benchmark for Compositional Text-to-video Generation (CVPR 2025 · similarity 0.82)
- The Devil is in the Prompts: Retrieval-Augmented Prompt Optimization for Text-to-Video Generation (CVPR 2025 · similarity 0.81)
- Weakly Supervised Video Scene Graph Generation via Natural Language Supervision (ICLR 2025 · similarity 0.81)
- Tag2Text: Guiding Vision-Language Model via Image Tagging (ICLR 2024 · similarity 0.81)
- SEINE: Short-to-Long Video Diffusion Model for Generative Transition and Prediction (ICLR 2024 · similarity 0.81)
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.
