Research map / Vision-language and multimodal / Human and embodied modeling
SEPS: Semantic-Enhanced Patch Slimming Framework for Fine-Grained Cross-Modal Alignment
Xinyu Mao, Junsi Li, Haoji Zhang, Yu Liang, Ming Sun
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 dense-text guidance alongside sparse text in patch selection (vs sparse only) (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 67.2 -> 80.5 | Improved |
| add the full consensus scoring rather than concatenating sparse and dense text (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 80.8 -> 86.9 | Improved |
| use weighted consensus instead of summing sparse and dense scores (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 81.1 -> 86.9 | Improved |
| add the adaptive aggregation network (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 83.7 -> 86.9 | Improved |
| add mean-value term to relevance-aware selection in SGMA (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 84.1 -> 86.9 | Improved |
| use salience-guided aggregation instead of mean pooling only (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 82.4 -> 86.9 | Improved |
| use dense text of the same image instead of another image from same dataset (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 79.3 -> 86.9 | Improved |
| use dense text from the same dataset rather than a different dataset (ablation) | SEPS · Flickr30K 1K | text-to-image R@1: 71.7 -> 79.3 | Improved |
| replace LAPS with SEPS on zero-shot COCO transfer | LAPS · MS-COCO 5K (zero-shot) | text-to-image R@1: 19.0 -> 44.3 | Improved |
| replace D2S-VSE with SEPS in online inference latency | D2S-VSE · Flickr30K inference | online inference latency: 78ms -> 107ms | Hurt |
| replace LAPS with SEPS on CLIP ViT-Base fine-tuning | LAPS · Flickr30K 1K (CLIP) | text-to-image R@1: 80.6 -> 93.1 | Improved |
| replace LAPS with SEPS on zero-shot RefCOCO grounding | LAPS · RefCOCO val | accuracy: 44.2 -> 48.7 | Improved |
Closest papers by meaning
- Linguistic-Aware Patch Slimming Framework for Fine-grained Cross-Modal Alignment (CVPR 2024 · similarity 0.94)
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.90)
- Beyond Global Similarity: Multi-Conditional Retrieval for Fine-Grained Cross-Modal Understanding (CVPR 2026 · similarity 0.89)
- Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models (NeurIPS 2025 · similarity 0.89)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · similarity 0.89)
- Patch Matters: Training-free Fine-grained Image Caption Enhancement via Local Perception (CVPR 2025 · 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.
