Research map / Vision-language and multimodal / Cross-modal learning
Seeing What You Miss: Vision-Language Pre-Training With Semantic Completion Learning
Yatai Ji, Rongcheng Tu, Jie Jiang, Weijie Kong, Chengfei Cai, Wenzhe Zhao, Hongfa Wang, Yujiu Yang, Wei Liu
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 MLM (ablation) | SCL pre-training · VQA2.0 | accuracy: 75.46 -> 77.59 | Improved |
| add VTM (ablation) | SCL pre-training · Flickr30K zero-shot | IR@1: 49.7 -> 70.4 | Improved |
| add contrastive learning (ablation) | SCL pre-training · Flickr30K zero-shot | IR@1: 67.8 -> 70.4 | Improved |
| add semantic completion learning (ablation) | SCL pre-training · Flickr30K zero-shot | IR@1: 67.0 -> 70.4 | Improved |
| add MVSC alongside MLSC (ablation) | SCL pre-training · Flickr30K zero-shot | IR@1: 67.0 -> 70.4 | Improved |
| add MLSC alongside MVSC (ablation) | SCL pre-training · Flickr30K zero-shot | IR@1: 69.5 -> 70.4 | Improved |
| set image mask ratio to 0.7 rather than 0.8, keeping text at 0.4 (ablation) | SCL · Flickr30K zero-shot | IR@1: 70.4 -> 68.8 | Hurt |
| set text mask ratio to 0.3 rather than 0.4, keeping image at 0.8 (ablation) | SCL · Flickr30K zero-shot | IR@1: 70.4 -> 69.5 | Hurt |
| set text mask ratio to 0.5 rather than 0.4, keeping image at 0.8 (ablation) | SCL · Flickr30K zero-shot | IR@1: 70.4 -> 68.8 | Hurt |
| set image mask ratio to 0.9 rather than 0.8, keeping text at 0.4 (ablation) | SCL · Flickr30K zero-shot | IR@1: 70.4 -> 68.6 | Hurt |
| replace Frame CLS temporal vision encoder with mean pooling (ablation) | video-text retrieval encoder · MSRVTT zero-shot video retrieval | R@1: 23.3 -> 23.0 | Hurt |
| replace Frame CLS temporal vision encoder with Global CLS (ablation) | video-text retrieval encoder · MSRVTT zero-shot video retrieval | R@1: 23.3 -> 22.5 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Clover — Clover: Towards a Unified Video-Language Alignment and Fusion Model (CVPR 2023) · better on 3 of 3 shared results
- BLIP — BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation (ICML 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Global and Local Semantic Completion Learning for Vision-Language Pre-Training (TPAMI 2025 · similarity 0.97)
- Masked Vision and Language Modeling for Multi-modal Representation Learning (ICLR 2023 · similarity 0.91)
- Leveraging per Image-Token Consistency for Vision-Language Pre-Training (CVPR 2023 · similarity 0.90)
- Single-Stream Multi-level Alignment for Vision-Language Pretraining (ECCV 2022 · similarity 0.90)
- VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set (NeurIPS 2025 · similarity 0.88)
- SyncMask: Synchronized Attentional Masking for Fashion-centric Vision-Language Pretraining (CVPR 2024 · similarity 0.87)
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.
