Research map / Vision-language and multimodal / Cross-modal learning
Global and Local Semantic Completion Learning for Vision-Language Pre-Training
Rong-Cheng Tu, Yatai Ji, 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 MGSC pre-training task (ablation) | MLM+VTM+CL baseline · Flickr30K zero-shot TR@1 | recall: 80.0 -> 86.2 | Improved |
| add MLM (ablation) | VTM+CL+MGSC · VQA2.0 test-dev | accuracy: 75.46 -> 77.59 | Improved |
| add VTM (ablation) | MLM+CL+MGSC · VQA2.0 test-dev | accuracy: 76.96 -> 77.59 | Improved |
| add CL (ablation) | MLM+VTM+MGSC · VQA2.0 test-dev | accuracy: 77.66 -> 77.59 | No clear effect |
| add MVSC (vision half of MGSC) (ablation) | MLM+VTM+CL · Flickr30K zero-shot TR@1 | recall: 80.0 -> 85.5 | Improved |
| add MLSC (language half of MGSC) (ablation) | MLM+VTM+CL · Flickr30K zero-shot TR@1 | recall: 80.0 -> 80.5 | Improved |
| add MLTC with image mask ratio 0.3 (ablation) | MGSC model · Flickr30K zero-shot IR@1 | recall: 70.4 -> 71.5 | Improved |
| add MLTC without MLM (ablation) | VTM+CL+MGSC · Flickr30K zero-shot IR@1 | recall: 66.9 -> 67.9 | Improved |
| add MLTC without MGSC (ablation) | VTM+CL+MLM · Flickr30K zero-shot IR@1 | recall: 66.4 -> 68.3 | Improved |
| replace InfoNCE with L2 loss in MGSC (ablation) | GLSCL · Flickr30K zero-shot IR@1 | recall: 71.5 -> 70.1 | Hurt |
| replace InfoNCE with L2 loss in MLTC (ablation) | GLSCL · Flickr30K zero-shot IR@1 | recall: 71.5 -> 69.6 | Hurt |
| raise image mask ratio from 0.8 to 0.9 (ablation) | MGSC · Flickr30K zero-shot IR@1 | recall: 70.4 -> 68.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Clover — Clover: Towards a Unified Video-Language Alignment and Fusion Model (CVPR 2023) · better on 3 of 3 shared results
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
- Seeing What You Miss: Vision-Language Pre-Training With Semantic Completion Learning (CVPR 2023 · similarity 0.97)
- Modality-Specialized Synergizers for Interleaved Vision-Language Generalists (ICLR 2025 · similarity 0.89)
- VL-SAE: Interpreting and Enhancing Vision-Language Alignment with a Unified Concept Set (NeurIPS 2025 · similarity 0.88)
- Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models (NeurIPS 2025 · similarity 0.88)
- Locality Alignment Improves Vision-Language Models (ICLR 2025 · similarity 0.87)
- Linguistic-Aware Patch Slimming Framework for Fine-grained Cross-Modal Alignment (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.
