Research map / Vision-language and multimodal / Human and embodied modeling
Unicoder-VL: A Universal Encoder for Vision and Language by Cross-Modal Pre-Training
Gen Li, Nan Duan, Yuejian Fang, Ming Gong, Daxin Jiang
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 |
|---|---|---|---|
| Pre-train on image-caption pairs and fine-tune for retrieval rather than train only on task-specific data (ablation) | Unicoder-VL · MSCOCO 1K test, sentence retrieval | R@1: 75.1 -> 84.3 | Improved |
| Pre-train on image-caption pairs and fine-tune for retrieval rather than train only on task-specific data (ablation) | Unicoder-VL · Flickr30k 1K test, sentence retrieval | R@1: 73.0 -> 86.2 | Improved |
| Pre-train on image-caption pairs and fine-tune for retrieval rather than train only on task-specific data (ablation) | Unicoder-VL · Flickr30k 1K test, image retrieval | R@1: 57.8 -> 71.5 | Improved |
| Use pretrained alignment scoring without task-specific fine-tuning | Unicoder-VL · Flickr30k 1K test, sentence retrieval | R@1: 64.3 -> 86.2 | Hurt |
| Increase the encoder depth from 6 to 12 layers (ablation) | Unicoder-VL · Flickr30k 1K test, sentence retrieval | R@1: 72.4 -> 86.2 | Improved |
| Increase the encoder depth from 12 to 24 layers (ablation) | Unicoder-VL · Flickr30k 1K test, image retrieval | R@1: 71.5 -> 73.6 | Improved |
| Increase pre-training data from 0 to 3M image-text pairs (ablation) | Unicoder-VL · Flickr30k 1K test, sentence retrieval | R@1: 73.0 -> 82.3 | Improved |
| Increase pre-training data from 3M to 3.8M image-text pairs (ablation) | Unicoder-VL · Flickr30k 1K test, image retrieval | R@1: 68.3 -> 71.5 | Improved |
| Use cross-modal pre-training for VCR instead of no pre-training (ablation) | Unicoder-VL · VCR val, Q→A | accuracy: 71.6 -> 72.6 | Improved |
| Use cross-modal pre-training for VCR instead of no pre-training (ablation) | Unicoder-VL · VCR val, Q→AR | accuracy: 52.3 -> 54.5 | Improved |
| Use caption-based cross-modal pre-training for cognitive VCR reasoning (ablation) | Unicoder-VL · VCR val, Q→A | accuracy: 71.6 -> 72.6 | No clear effect |
Closest papers by meaning
- UNITER: UNiversal Image-TExt Representation Learning (ECCV 2020 · similarity 0.91)
- UC2: Universal Cross-Lingual Cross-Modal Vision-and-Language Pre-Training (CVPR 2021 · similarity 0.90)
- Unified Vision-Language Pre-Training for Image Captioning and VQA (AAAI 2020 · similarity 0.90)
- Probing Inter-modality: Visual Parsing with Self-Attention for Vision-and-Language Pre-training (NeurIPS 2021 · similarity 0.90)
- Vision‑Language‑Vision Auto‑Encoder: Scalable Knowledge Distillation from Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Unveiling Encoder-Free Vision-Language Models (NeurIPS 2024 · similarity 0.89)
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.
