AccessLens: Auto-detecting Inaccessibility of Everyday Objects

Title of the Paper

AccessLens: Auto-detecting Inaccessibility of Everyday Objects

Paper Information

• Domain: Human-Computer Interaction, Accessibility Design, Computer Vision
• Keywords: Accessibility Design, 3D Printing Assistance, Object Detection, Human-Computer Interaction, Accessible Computing, Dataset, Indoor Scenes

Research Background and Problem Statement

• Problems or Challenges:

  1. Many everyday objects (e.g., door handles, switches) may pose usability challenges for certain individuals (e.g., those with physical or sensory impairments) in specific contexts.
  2. Traditional solutions for addressing accessibility issues, such as physical modifications or installing assistive devices, are often costly and fail to adapt to dynamic or contextual needs.
  3. Ordinary users (especially those without firsthand experience of disabilities) often lack awareness of accessibility issues, further delaying the promotion of community-wide accessibility awareness.

• Significance: Raising public awareness of hidden accessibility issues in daily environments can promote more inclusive design, realizing the principle of "design for one, benefit all."

• Research Motivation and Related Work:

  1. Existing datasets (e.g., ADE20K) lack annotations for object types and interaction contexts related to accessibility.
  2. Current systems often focus on specific disability types (e.g., wheelchair users) or specific scenarios, failing to capture temporary or contextual accessibility barriers.
  3. There is a lack of tools that can automatically identify accessibility issues and suggest low-cost solutions while enhancing users' awareness of accessibility.

Solution

• Method or Solution: The authors propose an end-to-end system, AccessLens, which includes the following three key modules:

  1. AccessDB/AccessReal Dataset: Contains over 10,000 re-annotated indoor scene objects for training and testing models that detect inaccessible objects, covering 21 accessibility categories.
  2. AccessMeta Metadata: Provides a semantic dictionary through 3D printing-assisted design, linking object interaction attributes with accessibility categories.
  3. User Application Toolkit: Through a mobile interface, users can scan indoor photos, and the system automatically detects accessibility issues and suggests targeted designs.

• Innovations:

  1. The first system to combine inaccessible object detection with 3D printing-assisted design recommendations.
  2. Creation of new datasets (AccessDB and AccessReal) specifically annotated for accessibility issues.
  3. Designed for users with limited accessibility awareness, providing plug-and-play 3D printing designs to lower technical and psychological barriers.
  4. Emphasizes that accessibility is a universal issue, not limited to individuals with diagnosed disabilities.

• Implementation Steps and Key Technologies:

  1. Data Construction: Re-annotated the ADE20K dataset to create AccessDB and collected modern indoor scenes to generate AccessReal.
  2. Training and Detection: Trained an inaccessible object detection model using RetinaNet and AccessDB.
  3. Metadata Design: Defined three main categories (action constraints, identification indicators, operational aids) and multifunctional 3D printing design semantic metadata.
  4. System Evaluation and User Feedback: Evaluated the system's accuracy and user acceptance through user studies and technical experiments.

Research Outcomes

• Specific Outcomes:

  1. AccessDB and AccessReal: Provided detailed annotated data for 21 accessibility categories, covering 10,467 object instances.
  2. AccessMeta: Created a metadata dictionary linking 280 types of 3D printing-assisted designs with 52 types of everyday objects.
  3. AccessLens User Interface and System Performance: Experiments demonstrated the system's ability to enhance users' awareness of accessibility issues, with the detector performing well in identifying inaccessible objects in modern scenes (mAP close to 15%).

• Advantages over Existing Solutions:

  1. More comprehensively addresses everyday accessibility issues, supporting contextual and situational accessibility awareness.
  2. Provides immediate, low-cost 3D printing solutions with actionable recommendations for users.
  3. Designed for ordinary users with no prior experience in assistive technologies, significantly lowering the barrier to use.

• Experimental or Evaluation Results:

  1. Achieved good detection performance on the AccessReal dataset, with the model capable of identifying small inaccessible objects in images.
  2. In user studies, AccessLens significantly improved users' ability to identify inaccessible objects and find solutions compared to written guidelines.
  3. Users generally found the system's recommended 3D printing designs easy to implement and cost-effective.

• Limitations and Future Directions:

  1. Limitations:
     • The current detection model has limitations in high-resolution scenarios.
     • The system's recommendations may encounter design conflicts (e.g., conflicting needs of multiple users).
     • Lacks automated customization of 3D models and comprehensive constraint detection.
  2. Future Directions:
     • Expand system functionality to accommodate more accessibility categories and real-world user scenarios.
     • Integrate customization features to meet more complex 3D printing needs.
     • Promote community collaboration to continuously optimize AccessMeta classification and expand the dataset through user feedback.
     • Combine automated tools and machine learning models to improve the accuracy of accessibility assessments and solution recommendations.