TL;DR
Researchers have introduced the Reachspace Database, a novel collection of over 10,000 first-person images capturing "reach-relevant" environments. This database aims to fill a critical gap in visual perception research by providing a comprehensive resource for studying how we process and interact with our immediate surroundings, which has been largely overlooked compared to object or scene perception.
What Problem Does This Paper Address?
Our daily lives are filled with interactions in "near-scale" environments – the spaces within arm's reach where we perform tasks like typing, cooking, or assembling objects. While research has extensively explored how we perceive individual objects and navigate large-scale environments ("scenes"), the intermediate scale of these "reachspaces" has been comparatively neglected. This oversight is problematic because recent findings suggest that the perceptual processing of reachspaces might differ significantly from that of objects and scenes. However, a major bottleneck has been the lack of high-quality, diverse image datasets specifically designed for studying these environments. Existing databases often focus on single objects or broad scene categories, failing to capture the rich, task-oriented visual input characteristic of reachspaces. This paper addresses this gap by introducing the Reachspace Database, a resource intended to accelerate research into the perceptual, cognitive, and neural processes underlying our understanding of the near-scale world.
Methods
The creation of the Reachspace Database involved a systematic, multi-stage approach:
- Category Generation: To overcome the lack of pre-existing lists of reach-relevant environments, the researchers adopted an "activity-centered" approach. They compiled lists of common hobbies, professions, and household chores, and then brainstormed specific tasks and activities associated with various scenes (e.g., "coffeeshop") and objects (e.g., "wrench"). Activities were included only if they involved near-space interactions or hand-object interactions within a stable, reachable-scale environment.
- Image Collection: Online image searches were conducted using keywords derived from the generated activity lists, often augmented with terms like "my" (e.g., "my crafting desk") or "setup" to encourage first-person perspectives and appropriate framing. Searches were performed across multiple search engines and languages to ensure cultural diversity. Some images were also original photographs taken by the researchers.
- Taxonomy Development: A four-part naming convention was developed to systematically categorize the images. This taxonomy includes:
- Setting: The broader location type (e.g., home, hotel, office).
- Room: The specific room or site within the setting (e.g., kitchen, conference room, campsite).
- Interaction Locus: The primary structure supporting interaction (e.g., table, countertop, control panel).
-
Action: The specific task afforded by the reachspace (e.g., eating, working, shopping). In some cases, a fifth level was added for finer distinctions (e.g., distinguishing between different types of restaurant place settings).
-
Quality Control: Images were included if they were real-world photographs, in full color, and met specific resolution criteria (>= 125x125 pixels). Crucially, images had to depict a view similar to what an observer would experience while actively engaging with the space, with the center of interaction roughly centered in the frame. Images showing clear views of people, hands, or faces, or those with significant watermarks or filter effects, were excluded.
Datasets
The paper introduces the Reachspace Database, a newly created benchmark.
- Reachspace Database: Contains over 11,276 images drawn from 351 categories (as of October 31, 2020). The database is available for download on the Open Science Framework (osf.io/bfyxk/) and can be previewed at www.reachspacedatabase.com.
- Existing Databases: The authors note that while databases for object perception (e.g., ImageNet, Things) and scene perception (e.g., Places) exist, none specifically focus on the reach-relevant scale.
Key Insights and Ideas
- Reachspaces are a distinct perceptual domain: The structural and visual differences between reachspaces, scenes, and objects suggest that they are processed differently by the visual system. This necessitates dedicated research and resources for this scale.
- Task-activity drives reachspace structure: The taxonomy and image collection process highlight a strong connection between the activities performed in a space and the space's identity. This suggests that understanding the "affordances" of a space is key to its perception.
- Specialization vs. Generality: While many reachspaces are highly specialized for a single activity (e.g., an airplane cockpit for piloting), a few "hub" locations (like tables or kitchens) support a wide range of diverse actions, indicating a non-uniform distribution of functional complexity.
- Taxonomy's Generative Power: The developed four-part taxonomy proved highly effective not only for organizing the database but also for discovering new reachspace categories by systematically varying its components.
- Natural Language Ambiguity: The relative vagueness of natural language labels for reachspaces (e.g., "desk") compared to the precise, multi-level taxonomy used in the database raises interesting questions about how we mentally represent and communicate about these environments.
Key Results
- The Reachspace Database contains 11,276 images across 351 categories.
- On average, there are 32 images per category, with a minimum of 10 images per category.
- The database encompasses 38 unique settings, 131 unique rooms, 161 unique interaction loci, and 143 unique actions.
- Analysis of the taxonomy reveals that 90.8% of setting-room-locus chains were associated with only 1 action, indicating that these factors strongly dictate the space's function. However, a tail exists where some chains support 3 or more actions.
- 50.0% of setting labels and 88.5% of room labels appeared only one to three times, suggesting high specificity in many environments. Conversely, a smaller percentage of settings and rooms were associated with a large number of reachspaces.
Limitations and Future Work
- Sampling Bias: The database is biased towards spaces that are commonly photographed and posted online, potentially underrepresenting professions with less visually documented workspaces (e.g., industrial kitchens, specialized machinery). The cultural context is also largely industrialized and Western.
- Canonical Names: The category names are descriptive labels developed by the researchers and may not reflect how people commonly refer to these spaces in everyday language. Further validation of these labels with broader user groups is suggested.
- Limited Scope of Labels: The four-level taxonomy, while comprehensive, cannot capture the full richness of language used to describe these spaces. For instance, a "kitchen counter" might afford "chopping," but also "slicing" and "dicing," which are not explicitly differentiated in the primary labels.
- Future Work: The authors suggest future versions could benefit from a fuller sample of related terms and broader validation of category labels. They also highlight ongoing work to measure similarity structures among reachspace images using human behavioral judgments and data-driven methods to uncover underlying perceptual dimensions. The ultimate goal is to encourage research that integrates reachspaces with navigable-scale scenes to build more complete models of human perception across all scales.
Summary updated