Research map / Vision-language and multimodal / Human and embodied modeling
Understanding Bias in Large-Scale Visual Datasets
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 a transformation-based framework (semantic, structural, color, frequency, synthetic) plus object-level and LLM-based language analysis to identify concrete visual forms of bias among YFCC, CC, and DataComp.
Problem
Prior work showed large-scale visual datasets can be easily classified by their origin, but the concrete visual attributes that make these datasets distinguishable remain unclear.
Approach
Isolate specific information types (semantics, structure, boundaries, color, frequency) via transformations of images, train a dataset classifier on each transformed set to measure how much bias each carries, then interpret semantic bias with object-level queries, topic models, and LLM summaries.
Key results
- Dataset classification on original images reaches 82.0% reference accuracy, and semantic segmentation maps reach 69.8%, while SAM contours reach 73.2% and depth 73.1%.
- Pixel shuffling drops accuracy to 52.2% (random order), while patch shuffling with patch size 16 nearly recovers the reference accuracy.
- Unconditionally generated DiT images are classified at 77.6% accuracy, indicating synthetic images inherit dataset bias.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Transform images into semantic segmentation masks | YCD dataset classification · YCD | accuracy: 82.0% -> 69.8% | Hurt |
| Transform images into object bounding boxes | YCD dataset classification · YCD | accuracy: 82.0% -> 61.9% | Hurt |
| Use long instead of short image captions | Caption-based YCD dataset classification · YCD | accuracy: 63.8% -> 66.1% | Helped |
| Reconstruct images with a pretrained VAE | YCD dataset classification · YCD | accuracy: 82.0% -> 77.4% | Hurt |
| Use SAM object contours instead of Canny edges | YCD dataset classification · YCD | accuracy: 71.0% -> 73.2% | Helped |
| Shuffle pixel positions | YCD dataset classification · YCD | accuracy: 82.0% -> 52.2% (random order); 82.0% -> 58.5% (fixed order) | Hurt |
| Keep only each image's mean RGB color | YCD dataset classification · YCD | accuracy: 48.5% | Hurt |
| Keep high-frequency image components instead of low-frequency components | YCD dataset classification · YCD | accuracy: 70.4% -> 79.2% | Helped |
| Generate images unconditionally with dataset-specific DiTs | YCD dataset classification · Synthetic images | accuracy: 77.6% | Helped |
| Use ViT-Base instead of ViT-Large for SAM contour extraction (ablation) | YCD dataset classification on SAM contours · YCD | accuracy: 72.3% vs 73.1% | No clear effect |
| Represent segmentation masks as a 150-channel binary mask instead of a color-coded RGB image (ablation) | YCD dataset classification on semantic masks · YCD | accuracy: 69.8% -> 67.6% | Hurt |
| Combine pixel shuffling with SAM contours | YCD dataset classification · YCD | accuracy: 52.2% and 73.1% -> 74.2% | Helped |
Surprising findings
- Patch shuffling with patch size 16 almost fully recovers classification accuracy, showing local patch structure suffices to identify datasets.
- Mean RGB values alone yield 48.5% accuracy, showing color statistics carry non-trivial dataset signal.
Limitations
- Analysis limited to three datasets (YFCC, CC, DataComp).
- Object-level analysis restricted to closed object vocabularies from ImageNet, LVIS, and ADE20K.
- Reference accuracy is lower than original work due to shorter training (30 epochs).
Experimental setup
Datasets: YFCC100M CC12M DataComp-1B ADE20K LVIS ImageNet-1K
Benchmarks: YFCC vs CC vs DataComp dataset classification accuracy
Metrics: validation accuracy standard deviation over three repeats
Baselines: VisDiff Liu and He original dataset classification
Builds on: Name That Dataset (Torralba and Efros) Liu and He dataset classification VisDiff DataComp
Closest papers by meaning
- Classifier-to-Bias: Toward Unsupervised Automatic Bias Detection for Visual Classifiers (CVPR 2025 · similarity 0.89)
- ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts (NeurIPS 2025 · similarity 0.88)
- Common Sense Bias Modeling for Classification Tasks (AAAI 2025 · similarity 0.88)
- A Decade's Battle on Dataset Bias: Are We There Yet? (ICLR 2025 · similarity 0.88)
- Discovering and Mitigating Visual Biases through Keyword Explanation (CVPR 2024 · similarity 0.88)
- Visual Data Diagnosis and Debiasing with Concept Graphs (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.