Research map / Vision-language and multimodal / Cross-modal learning
Discrete-Continuous Action Space Policy Gradient-Based Attention for Image-Text Matching
Shiyang Yan, Li Yu, Yuan Xie
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 discrete-continuous PG-based attention | Flickr30k matching baseline · Flickr30k | Caption retrieval R@1: 70.9 -> 81.0 | Improved |
| add Instance Loss to Triplet Loss (ablation) | Flickr30k matching model · Flickr30k | Caption retrieval R@1: 68.2 -> 69.3 | Improved |
| add Text Decoding Loss to Triplet + Instance (ablation) | Flickr30k matching model · Flickr30k | Caption retrieval R@1: 69.3 -> 70.9 | Improved |
| use discrete-continuous PG instead of discrete-only PG (ablation) | Flickr30k matching model · Flickr30k | Caption retrieval R@1: 78.0 -> 81.0 | Improved |
| use discrete-continuous PG instead of single-Gaussian continuous PG (ablation) | Flickr30k matching model · Flickr30k | Caption retrieval R@1: 76.8 -> 81.0 | Improved |
| use combined R@1 and AP reward instead of R@1-only reward (ablation) | Flickr30k discrete-continuous PG · Flickr30k | Caption retrieval R@1: 79.3 -> 81.0 | Improved |
| use combined R@1 and AP reward instead of AP-only reward (ablation) | Flickr30k discrete-continuous PG · Flickr30k | Caption retrieval R@1: 80.4 -> 81.0 | Improved |
| set λ to 20 rather than 10 (ablation) | Flickr30k discrete-continuous PG · Flickr30k | Caption retrieval R@1: 80.2 -> 81.0 | Improved |
| set λ to 20 rather than 30 (ablation) | Flickr30k discrete-continuous PG · Flickr30k | Caption retrieval R@1: 79.2 -> 81.0 | Improved |
| add a PG baseline (ablation) | Flickr30k discrete-continuous PG · Flickr30k | Caption retrieval R@1: 81.0 -> 80.3 | Hurt |
| add a multi-head mechanism (ablation) | Flickr30k discrete-continuous PG · Flickr30k | Caption retrieval R@1: 81.0 -> 81.4 | Improved |
| use pre-trained GloVe word embeddings (ablation) | Flickr30k discrete-continuous PG · Flickr30k | Caption retrieval R@1: 81.0 -> 82.8 | Improved |
Closest papers by meaning
- Identity-Aware Textual-Visual Matching With Latent Co-Attention (ICCV 2017 · similarity 0.87)
- IMRAM: Iterative Matching With Recurrent Attention Memory for Cross-Modal Image-Text Retrieval (CVPR 2020 · similarity 0.86)
- Negative-Aware Attention Framework for Image-Text Matching (CVPR 2022 · similarity 0.86)
- Stacked Cross Attention for Image-Text Matching (ECCV 2018 · similarity 0.86)
- Text Embedding is Not All You Need: Attention Control for Text-to-Image Semantic Alignment with Text Self-Attention Maps (CVPR 2025 · similarity 0.85)
- CoMat: Aligning Text-to-Image Diffusion Model with Image-to-Text Concept Matching (NeurIPS 2024 · similarity 0.84)
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.
