Abstract: DATA REPORT article Front. Psychol., 23 May 2016 Sec. Perception Science Volume 7 - 2016 | https://doi.org/10.3389/fpsyg.2016.00749
Introduction: The Need for Controlled Stimuli
In visual search experiments, researchers frequently utilize photographs to study how humans locate objects. However, standard image sets—such as stock photography or advertisements—are often riddled with "artifacts" that bias experimental results. Common issues include central viewing bias (where photographers center the subject) and artificial depth-of-field cues that do not reflect natural vision. Furthermore, digitally "pasting" objects into scenes via software like Photoshop can create inconsistencies in lighting and shading that the human visual system detects instantly.
The Berlin Object in Scene (BOiS) database addresses these challenges by providing a "clean" dataset where photographic parameters are strictly controlled, objects are physically present during photography to ensure consistent lighting, and contextual expectations are mathematically quantified.
Methodology: Physical Realism and Contextual Priors
The researchers employed a rigorous, tripod-based photographic process to ensure consistency across 130 distinct scenes. For each scene, three identical shots were captured:
- Expected Location: The target object is placed in a typical setting (e.g., a watering can on the grass).
- Unexpected Location: The target object is placed in an atypical setting (e.g., a watering can on a roof).
- Absent: The scene is photographed without the target object.
To move beyond subjective intuition regarding object placement, the authors developed Contextual Prior Mapping. Ten naive subjects segmented the "target absent" images, marking regions based on four levels of likelihood for finding a specific object. These responses were averaged into grayscale heatmaps representing the "contextual prior," allowing researchers to distinguish between "bottom-up" saliency (e.g., a bright color) and "top-down" expectations.
Dataset Composition
The BOiS database includes a comprehensive suite of high-resolution ($4272 \times 2848$) stimuli:
- 390 Scene Images: 130 scenes across three versions (expected, unexpected, absent).
- 30,160 Object Images: 240 different views of each target object against a black background, providing "pre-search" visual cues.
- Diverse Environments: 107 indoor scenes (offices, hospitals, kitchens) and 23 outdoor scenes (parks, gardens).
- Rich Metadata: Includes segmentation masks, saliency measures (using the Itti & Koch model), and quantified contextual prior values.
Key Insights and Validation
A primary strength of BOiS is its physical realism. Approximately 75% of the scenes were photographed exactly as found, while 25% were manually rearranged in situ to ensure that shadows, reflections, and lighting remained physically consistent. Additionally, the authors intentionally kept central areas and outer margins clear of target objects, mitigating central viewing bias and allowing researchers to crop images without losing the target.
Statistical validation confirmed the effectiveness of the dataset's design. A paired t-test showed a massive statistical difference ($p < 10^{-38}$) between the contextual prior values of "expected" versus "unexpected" locations. In 129 out of 130 scenes, the authors' subjective placement matched the high-probability regions identified by independent subjects.
Limitations and Future Directions
While BOiS is highly optimized for controlled psychophysical experiments, its scale (130 scenes) is relatively small compared to massive computer vision datasets like COCO or ImageNet used for deep learning. Furthermore, the database currently consists of static images; future iterations may incorporate dynamic video to study visual search in moving environments. Researchers are also advised that in one specific instance (Scene 118), human prior maps did not align with the initial "expected/unexpected" categorization, and it may be excluded depending on the study's requirements.
Summary updated