Research map / Label-efficient and robust learning / Domain adaptation and generalization

Bridge Then Begin Anew: Generating Target-Relevant Intermediate Model for Source-Free Visual Emotion Adaptation

AAAI 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Bridge then Begin Anew (BBA), a two-step source-free domain adaptation framework for visual emotion recognition that first generates an intermediate bridge model and then trains a target model from scratch with polarity-constrained losses.

Problem

Source-free domain adaptation for visual emotion recognition, where labeled source emotion data is unavailable during adaptation and large affective gaps between VER datasets yield low-confidence pseudo-labels and source-model overfitting.

Approach

Generate a domain-bridged intermediate model (via clustering-based pseudo-label refinement, masked-image consistency, and EMA source teacher) and then train a randomly initialized target model from scratch guided by the bridge model, with an emotion polarity loss added to self-labeling and information-maximization objectives.

Key results

• On FI→EmoSet-style settings, BBA reaches 57.36% accuracy on EmoSet→FI versus 55.25% for the best prior SFDA baseline (DINE).
• On Emotion6→FI, BBA reaches 78.12% accuracy versus 72.42% for TPDS.
• The paper reports an average improvement of +3.03 over state-of-the-art SFDA methods across six settings.

Measured effects reported in the paper

Surprising findings

• Standard classification models adapted from source tend to outperform VER models by a large margin, indicating VER datasets lack explicit class characterization.

Limitations

• Evaluated only on image-based VER datasets and a limited set of six adaptation settings.
• Emotion6 is treated as binary classification in cross-domain settings, which simplifies the task.

Experimental setup

Datasets: EmoSet FI Emotion6 ArtPhoto Office-Home

Benchmarks: EmoSet FI Emotion6 ArtPhoto

Metrics: accuracy macro precision macro recall macro F1 weighted precision weighted recall weighted F1

Baselines: Source only SHOT SHOT++ G-SFDA AaD DaC DINE TPDS MCC ELS MIC Oracle

Builds on: SHOT self-distillation information maximization

Closest papers by meaning

• Knowledge-Aligned Counterfactual-Enhancement Diffusion Perception for Unsupervised Cross-Domain Visual Emotion Recognition (CVPR 2025 · similarity 0.89)
• Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach (ICLR 2026 · similarity 0.88)
• AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.87)
• VidEmo: Affective-Tree Reasoning for Emotion-Centric Video Foundation Models (NeurIPS 2025 · similarity 0.86)
• $E^3$: Exploring Embodied Emotion Through A Large-Scale Egocentric Video Dataset (NeurIPS 2024 · similarity 0.86)
• Dual-View Learning for Conversational Emotion Recognition Through Context and Emotion-Shift Modeling (AAAI 2025 · similarity 0.86)