"It's trained by non-disabled people": Evaluating How Image Quality Affects Product Captioning with Vision-Language Models

1. Problem Statement

Vision-Language Models (VLMs) are increasingly deployed to assist Blind and Low-Vision (BLV) individuals in identifying everyday products (e.g., food, medication, household goods). However, a critical gap exists in understanding how real-world image quality issues—specifically blur, misframing, and rotation, which are common when BLV users take photos—affect the accuracy of VLM-generated captions.

Current evaluation methods often rely on high-quality datasets (e.g., ImageNet, MS COCO) or synthetic distortions, failing to capture the specific nuances of images taken by BLV users. Furthermore, standard captioning metrics (e.g., BLEU, CIDEr) measure text similarity rather than factual correctness, leading to "false positives" where a caption sounds reasonable but contains dangerous errors (e.g., misidentifying a medication or allergen). The paper argues that without addressing these quality issues, VLMs remain unreliable for high-stakes tasks like identifying food ingredients or medication dosages.

2. Methodology

The study employs a two-part mixed-method approach:

Study 1: User Survey (Qualitative & Quantitative)

• Participants: 86 BLV individuals who regularly use AI tools for image captioning.
• Focus: Investigated user preferences between AI and human assistance, challenges in taking photos, perceived impact of image quality on caption accuracy, and the frequency of specific error types.
• Key Findings from Survey:
  • Privacy vs. Accuracy: Users prefer AI for privacy-sensitive tasks (e.g., reading food labels) but rely on humans for safety-critical tasks (e.g., medication).
  • Photo Taking: Taking a "good" photo is the hardest part of the process; 67% of users found it the most difficult step.
  • Error Types: The most frequent and dangerous error is missing critical information (e.g., brand name, specific variety/ingredients) rather than complete hallucinations.
  • Feedback Gap: Current tools provide insufficient guidance on how to fix image quality issues (e.g., telling a user an image is blurry but not how to stabilize the hand).

Study 2: Systematic Evaluation (Quantitative)

• Dataset Construction: The authors created a new annotated dataset of 1,859 product images taken by BLV people, derived from the VizWiz dataset.
  • High-Quality Subset: 729 images with no significant quality issues.
  • Low-Quality Subset: 1,130 images with specific issues: Blur, Framing (misframed), and Rotation.
  • Annotation: Human researchers manually verified product details (Type, Brand, Variety) and VLM outputs, creating a "ground truth" that standard metrics cannot provide.
• Models Evaluated: Four state-of-the-art VLMs:
  1. GPT-4.1 (Closed-source, powers Seeing AI/Be My AI).
  2. Gemini 2.5 Flash (Closed-source).
  3. Llama 3.2 90B (Open-source).
  4. Molmo 72B (Open-source).
• Analysis: Logistic regression was used to model the relationship between image quality issues, product properties (e.g., rounded labels, text panels), and identification accuracy.

3. Key Results

Performance Degradation

• High-Quality Images: All models performed well, with GPT-4.1 achieving 98.5% accuracy and Gemini 95.7%. Open-source models (Llama, Molmo) were slightly lower (~86%).
• Low-Quality Images: Accuracy dropped significantly across all models.
  • GPT-4.1: Dropped to 74.9%.
  • Gemini: Dropped to 71.7%.
  • Llama: Dropped to 44.1%.
  • Molmo: Dropped to 36.1%.
• Compound Issues: When multiple issues (e.g., blur + rotation) occurred simultaneously, performance worsened further. GPT-4.1 dropped to 69.4%, while Molmo fell to 27.8%.

Specific Vulnerabilities

• Blur: The most detrimental factor, reducing the odds of correct identification by 88.3%. Models struggle to read fine text (flavors, ingredients) on blurred labels.
• Framing: The second most problematic issue (84.5% reduction in odds). Models often fail to infer missing text or brand names when the product is partially out of frame.
• Rotation: Less severe than blur/framing but still significant (79.5% reduction). Models struggle to orient text and identify products when rotated >45 degrees.
• Product Properties:
  • Rounded Labels: Generally handled well by top models unless combined with other issues.
  • Text Panels: Large text blocks (e.g., nutrition facts) sometimes caused models to over-focus on text and ignore visual context, leading to hallucinations (e.g., misidentifying "Organics" as a brand name).

Model-Specific Insights

• GPT-4.1: Most robust overall, particularly against blur and rotation.
• Gemini: Showed slightly better resistance to blur than GPT but was more sensitive to framing issues.
• Molmo: Most sensitive to rotation and framing; struggled significantly with text-heavy panels.
• Llama: Highly sensitive to blur.

4. Key Contributions

1. Empirical Evidence of the "Quality Gap": Demonstrated that VLMs, while excellent on clean data, suffer severe accuracy drops (up to 60% for open-source models) on the low-quality images typical of BLV users.
2. Disability-Centered Evaluation Framework: Proposed a methodology that moves beyond standard text-matching metrics. The study emphasizes human-in-the-loop verification of specific product attributes (Brand, Variety) to detect "false positives" where the caption is plausible but factually wrong.
3. Annotated Dataset: Released a curated dataset of 1,859 BLV-taken product images with detailed ground-truth annotations for product type, brand, and variety, specifically targeting quality issues.
4. Actionable Recommendations: Provided concrete directions for the HCI and ML communities to improve VLM reliability.

5. Significance and Recommendations

The paper argues that the current paradigm of asking BLV users to "take better photos" is insufficient and places an undue burden on the user. Instead, the authors propose a disability-centered approach to model development:

• Data Curation: Training data must include systematic degradations (blur, rotation) representative of real-world BLV photography, not just high-quality images.
• Model Training:
  • Open-Source Models: Prioritize fine-tuning open-source models (like Llama/Molmo) on synthetic degraded data to close the gap with closed-source models while preserving user privacy.
  • Loss Functions: Develop loss functions that penalize missing critical details (brand/variety) rather than just semantic similarity.
• Inference-Time Techniques:
  • Image Reconstruction: Use inpainting or super-resolution to repair images before captioning.
  • Partial Abstention: Instead of refusing to answer, models should identify which parts of the image they are confident about and explicitly state uncertainty for the rest (e.g., "I see a can of soup, but the brand is too blurry to read").
• User Feedback: Design tools that provide real-time, actionable feedback on how to fix a photo (e.g., "Rotate the can to the right") rather than just stating the image is "bad."

Conclusion: The paper concludes that for VLMs to be truly accessible and safe for BLV users, the research community must shift from optimizing for "average" image quality to optimizing for the worst-case scenarios inherent in the daily lives of disabled users.