Research map / Vision-language and multimodal / Human and embodied modeling
Contemplating Visual Emotions: Understanding and Overcoming Dataset Bias
Rameswar Panda, Jianming Zhang, Haoxiang Li, Joon-Young Lee, Xin Lu, Amit K. Roy-Chowdhury
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 |
|---|---|---|---|
| train on large web WEBEmo data instead of Deep Emotion | ResNet-50 · Binary cross-dataset (mean others) | accuracy (%): 63.52 -> 72.76 | Improved |
| train on WEBEmo instead of Emotion-6, tested on Emotion-6 | ResNet-50 · Emotion-6 test | accuracy (%): 77.72 -> 78.38 | Improved |
| add curriculum (2-6-25) instead of direct learning (ablation) | ResNet-50 · Exploration study (mean others) | accuracy (%): 68.13 -> 72.76 | Improved |
| add curriculum instead of self-directed learning (ablation) | ResNet-50 · Exploration study (mean others) | accuracy (%): 69.82 -> 72.76 | Improved |
| add curriculum instead of joint learning (ablation) | ResNet-50 · Exploration study (mean others) | accuracy (%): 70.38 -> 72.76 | Improved |
| use three-stage curriculum 2-6-25 instead of two-stage 2-6 (ablation) | ResNet-50 · Self test and mean others | accuracy (%): 78.21 -> 81.41 self; 70.05 -> 72.76 others | Improved |
| use 25 emotion categories instead of six for direct learning (ablation) | ResNet-50 · Direct learning self test | accuracy (%): 2.31% drop with six categories | Improved |
| train on WEBEmo versus Deep Emotion (sadness negative bias) | ResNet-50 · sadness vs non-sadness | % drop self to others: 25.09 -> -0.73 | Improved |
| use WEBEmo curriculum features instead of ImageNet features | ResNet-50 · UnBiasedEmo | accuracy (%): 64.20 -> 74.27 | Improved |
| use WEBEmo curriculum features instead of Deep Emotion features | ResNet-50 · UnBiasedEmo | accuracy (%): 68.88 -> 74.27 | Improved |
| use curriculum features instead of ImageNet features | ResNet-50 · Image Advertisement | accuracy (%): 23.42 -> 27.96 | Improved |
| use curriculum features instead of joint learning (ablation) | ResNet-50 · VideoStory-P14 | accuracy (%): 47.25 -> 49.22 | Improved |
Closest papers by meaning
- Building a Large Scale Dataset for Image Emotion Recognition: The Fine Print and The Benchmark (AAAI 2016 · similarity 0.94)
- Learning Visual Emotion Representations From Web Data (CVPR 2020 · similarity 0.91)
- Bridge Then Begin Anew: Generating Target-Relevant Intermediate Model for Source-Free Visual Emotion Adaptation (AAAI 2025 · similarity 0.89)
- Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach (ICLR 2026 · similarity 0.88)
- FindingEmo: An Image Dataset for Emotion Recognition in the Wild (NeurIPS 2024 · similarity 0.87)
- To Err Like Human: Affective Bias-Inspired Measures for Visual Emotion Recognition Evaluation (NeurIPS 2024 · similarity 0.87)
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.
