Research map / Vision-language and multimodal / Cross-modal learning
Efficient and High-Fidelity Omni Modality Retrieval
Chuong Huynh, Manh Luong, Abhinav Shrivastava
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 |
|---|---|---|---|
| use ASWP pooling instead of single [EOS] vector (ablation) | OmniRet · ablation avg recall | average recall: 43.4 -> 50.2 | Improved |
| use ASWP instead of 16 multi-vector late interaction (ablation) | OmniRet · ablation avg recall | average recall: 49.8 -> 50.2 | No clear effect |
| use STM hard selection instead of average pooling in ASWP (ablation) | OmniRet · ablation avg recall | average recall: 20.7 -> 50.2 | Improved |
| use STM instead of max pooling or weighted sum (ablation) | OmniRet · ablation avg recall | average recall: 49.2 / 48.5 -> 50.2 | Improved |
| add the shared media resampler (ablation) | OmniRet · ablation avg recall | average recall: 46.7 -> 50.2 | Improved |
| add modality-specific media latents to the shared resampler (ablation) | OmniRet · ablation avg recall | average recall: 49.8 -> 50.2 | No clear effect |
| add diversity loss on resampled tokens (ablation) | OmniRet · ablation avg recall | average recall: 47.1 -> 50.2 | Improved |
| add triplet loss (ablation) | OmniRet · ablation avg recall | average recall: 49.7 -> 50.2 | No clear effect |
| increase projections L from 1536 to 4096 (ablation) | OmniRet · ablation avg recall | average recall: 47.5 -> 50.2 | Improved |
| use OmniRet rather than VLM2VecV2 on image classification | OmniRet · MMEBv2 Image CLS | Recall@1: 62.9 -> 51.7 | Hurt |
| use OmniRet rather than VLM2VecV2 on video tasks | OmniRet · MMEBv2 Video | Recall@1: CLS 39.3 -> 48.6; RET 28.8 -> 36.5 | Improved |
| use single-stage OmniRet vs caption+Gemma pipeline | OmniRet · ACM composed audio | Recall@5: 44.6 (pipeline) vs 23.0 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- CLAP — CLAP: Isolating Content from Style through Contrastive Learning with Augmented Prompts (ECCV 2024) · better on 2 of 2 shared results
- ImageBind — ImageBind: One Embedding Space To Bind Them All (CVPR 2023) · better on 0 of 1 shared result
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
- Omni-Interactive Universal Embedder (NeurIPS 2026 · similarity 0.91)
- Bridging Modalities: Improving Universal Multimodal Retrieval by Multimodal Large Language Models (CVPR 2025 · similarity 0.89)
- OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models (NeurIPS 2026 · similarity 0.89)
- MM-EMBED: UNIVERSAL MULTIMODAL RETRIEVAL WITH MULTIMODAL LLMS (ICLR 2025 · similarity 0.89)
- AVOC: Enhancing Hour-Level Audio-Video Understanding in Omni-Modal LLMs via Retrieval-Inspired Token Compression (NeurIPS 2026 · similarity 0.88)
- OmniBench: Towards The Future of Universal Omni-Language Models (NeurIPS 2025 · 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.
