InsightVQA
收藏资源简介:
InsightVQA是由华东师范大学构建的大规模分层视觉问答数据集,旨在推动视觉情感理解与认知推理的研究。该数据集包含从六个公开来源收集并经过严格筛选的138K高质量图像,并标注了总计725K个问答对,内容涵盖感知、理解与认知三个层次。数据构建过程采用了伪标签验证、专家审核及多阶段质量控制流程,通过视觉触发提取和约束引导生成等技术确保标注的准确性与丰富性。该数据集主要应用于多模态大语言模型的情感认知能力评估与训练,致力于解决传统情感识别任务中缺乏可解释性、因果推理及高层认知分析能力的问题。
InsightVQA is a large-scale hierarchical visual question answering (VQA) dataset developed by East China Normal University, which aims to promote research on visual emotion understanding and cognitive reasoning. It includes 138,000 high-quality images collected from six public sources and rigorously filtered, with a total of 725,000 annotated question-answer pairs covering three levels: perception, comprehension, and cognition. The dataset construction process adopts workflows such as pseudo-label verification, expert review, and multi-stage quality control, and ensures the accuracy and richness of annotations via technologies including visual trigger extraction and constraint-guided generation. This dataset is primarily used for the evaluation and training of multimodal large language models' emotional cognition capabilities, and endeavors to address the limitations of traditional emotion recognition tasks, such as the lack of interpretability, causal reasoning, and high-level cognitive analysis abilities.
数据集名称
InsightVQA: High-Dimensional Emotion-Cognitive Visual Question Answering Benchmark
数据集简介
InsightVQA是一个面向高维情感认知的层次化视觉问答数据集,旨在推动模型从情感识别、情感触发理解到认知推理的完整推理能力。
数据规模
- 图像总数:138,000张(从351,000张公开来源图像中经多阶段筛选得到)。
- 问答对总数:725,000个。
数据集构成
数据集包含三个层次化的QA任务:
- 感知(Perception):情感与效价识别,共276,000个标签与效价QA对。
- 理解(Understanding):基于视觉触发提取的视觉、上下文与反事实QA,共330,000个QA对。
- 认知(Cognition):意图预测与顺序洞察推理,共119,000个QA对。
数据来源
图像收集自六个公开来源,经严格多阶段过滤后得到138,000张高置信度图像。
评价基准
InsightVQA-Bench:包含30,000个样本的高质量评估基准,用于细粒度评估。
方法(InsightNet)
基于InsightVQA数据集,提出InsightNet,一种统一架构的情绪调优多模态大语言模型(MLLM)基线。采用LoRA微调策略在Qwen2.5-VL-7B上进行层级化感知、理解与认知的训练,并通过指令监督微调(SFT)增强复杂情感场景下的认知深度与泛化能力。
实验结果(部分)
在InsightVQA-Bench上,InsightNet在三个任务中均取得领先性能:
- 感知任务:ACC 76.25%,F1 90.56%
- 理解任务:Ranking 82.79%,Top-1 71.21%
- 认知任务:ACC 69.18%
与开源/闭源MLLM及情感专用模型相比,InsightNet在所有指标上表现最佳(对比模型包括GPT-4o、Gemini-2.5-flash、Claude-3.7-sonnet等)。

- 1InsightVQA: High-Dimensional Emotion-Cognitive Visual Question Answering Benchmark华东师范大学 · 2026年



