SSA-ME：显式主体建模解决多模态检索中的视觉忽视与语义漂移

This post introduces the SSA-ME framework, which addresses visual neglect and semantic alignment bias in multimodal retrieval using salient subject-aware modeling and feature regeneration modules. It achieves state-of-the-art (SOTA) performance on the MMEB benchmark. The following floors break down the problem background, method details, experimental results, and more.

Based on Large Multimodal Models (LMMs), unified multimodal retrieval (UMR) has made progress, but existing embedding methods rely on sample-level contrastive learning and ignore subject-level semantic modeling. This leads to two key issues:

1. Semantic Alignment Bias: When the model fails to accurately map text to visual subjects (e.g., focusing on red flowers instead of the red bird in the query).
2. Visual Modal Neglect: Over-reliance on text clues, underutilizing visual information even when relevant.

To solve the above problems, researchers propose the SSA-ME (Salient Subject-Aware Multimodal Embedding) framework. Its core idea is to explicitly model salient visual subjects to guide the model to better understand and use visual information, while improving cross-modal semantic alignment.

The SSA-ME framework consists of three key components:

1. Salient Subject Identification & Emphasis: Uses LMM (for semantic relevance) and visual expert models (for visual salience) to identify significant subjects, then emphasizes these regions (via weighted features and background suppression).
2. Salient-Guided Alignment Objectives: Beyond sample-level contrastive learning, it adds subject-level alignment (ensures text subjects match visual regions, punishes misalignment even if samples are 'matched').
3. Feature Regeneration Module: Recalibrates visual features using salience maps (weighted aggregation, noise suppression, semantic enhancement) to balance visual and text modalities.

SSA-ME was evaluated on the MMEB (Massive Multimodal Embedding Benchmark), a large-scale multimodal retrieval benchmark. Key results:

• Achieved state-of-the-art (SOTA) performance.
• Ablation studies confirmed: Explicit salience modeling outperforms implicit learning; subject-level alignment brings extra gains; feature regeneration improves modal balance.

Qualitative Analysis:

• Attention visualization shows SSA-ME focuses on relevant subjects (e.g., dog and grass for 'dog running on grass') and suppresses irrelevant elements.
• Error case: Baseline matches 'person in red' to red buildings, while SSA-ME correctly matches to people in red.

Technical Insights: Why subject-level modeling matters?

• Aligns with human semantic understanding (naturally focusing on subjects).
• Solves visual neglect via explicit visual constraints.
• Enhances model interpretability (check where the model 'looks').

Current Limitations:

1. Extra computational cost from salience detection.
2. Limited to visual and semantic salience (ignores other types).
3. Challenges in complex scenes with multiple interacting subjects.

Future Directions:

1. Develop lightweight salience detection methods.
2. Explore dynamic salience (context-dependent).
3. Extend to model subject relationships.
4. Apply to multilingual scenarios.

SSA-ME effectively solves visual neglect and semantic drift in multimodal retrieval via explicit salient subject modeling. Its salient-guided alignment and feature regeneration modules enable balanced, semantically accurate cross-modal learning. This work highlights a key principle: explicit subject-level modeling is essential for true multimodal fusion, which will be valuable for future multimodal AI applications.