Research map / Vision-language and multimodal / Cross-modal learning
Preserving Semantic Neighborhoods for Robust Cross-modal Retrieval
Christopher Thomas, Adriana Kovashka
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 within-modality text and image neighbor losses (L_text, L_img) | Ang+NP+Sym · Politics | Recall@1 image-to-text: 0.6270 -> 0.6467 | Improved |
| add within-modality text and image neighbor losses | Ang+NP+Sym · GoodNews | Recall@1 image-to-text: 0.8704 -> 0.8849 | Improved |
| add within-modality text and image neighbor losses | Ang+NP+Sym · Conceptual Captions | Recall@1 image-to-text: 0.7687 -> 0.7760 | Improved |
| add within-modality text and image neighbor losses | Ang+NP+Sym · COCO | Recall@1 image-to-text: 0.6976 -> 0.6900 | Hurt |
| add the proposed losses to PVSE backbone | PVSE · GoodNews | Recall@1 image-to-text: 0.8724 -> 0.8867 | Improved |
| add the proposed losses to PVSE backbone | PVSE · Flickr30K | Recall@1 image-to-text: 0.5539 -> 0.5559 | No clear effect |
| add the proposed losses to triplet-loss baseline | Trip+NP+Sym · Politics | Recall@1 image-to-text: 0.4742 -> 0.4940 | Improved |
| add the proposed losses to triplet-loss baseline | Trip+NP+Sym · COCO | Recall@1 image-to-text: 0.4957 -> 0.4790 | Hurt |
| compute neighbors in Doc2Vec text space instead of image feature space (ablation) | Ours · Politics | Recall@1 image-to-text: 0.6370 -> 0.6467 | Improved |
| add L_text (ablation) | Ours (Ang) · GoodNews | Recall@1 image-to-text: 0.8786 -> 0.8849 | Improved |
| add L_img (ablation) | Ours (Ang) · Politics | Recall@1 image-to-text: 0.6390 -> 0.6467 | Improved |
| add the proposed losses to preserve Doc2Vec neighborhoods in joint space | Ang+NP+Sym · Politics (text neighborhood preservation) | neighbor overlap (text): 0.1544 -> 0.1703 | Improved |
Closest papers by meaning
- Cross-Modal Guided Visual Representation Learning for Social Image Retrieval (TPAMI 2024 · similarity 0.91)
- NeighborRetr: Balancing Hub Centrality in Cross-Modal Retrieval (CVPR 2025 · similarity 0.90)
- Accept the Modality Gap: An Exploration in the Hyperbolic Space (CVPR 2024 · similarity 0.90)
- Cross the Gap: Exposing the Intra-modal Misalignment in CLIP via Modality Inversion (ICLR 2025 · similarity 0.89)
- WikiDO: A New Benchmark Evaluating Cross-Modal Retrieval for Vision-Language Models (NeurIPS 2024 · similarity 0.88)
- Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data (ICLR 2024 · similarity 0.88)
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.
