Bounding Boxes
Definition
A Bounding Box is a fundamental geometric primitive in computer-vision used to localize and define the spatial extent of an object within an image or video frame. It is typically represented by a tuple (x_min, y_min, x_max, y_max) or (center_x, center_y, width, height) defining a rectangular region.
Core Characteristics
- Rectangular Constraint: Standard bounding boxes are axis-aligned rectangles, limiting precise fitting for rotated or irregularly shaped objects.
- Granularity: Provides coarse localization compared to Semantic Segmentation or Instance Segmentation, which offer pixel-level precision.
- Efficiency: Computationally lightweight, enabling real-time inference in resource-constrained environments.
Applications
- Object Detection: Identifying and localizing instance objects in natural images.
- Document AI & OCR: Essential for extracting structured data from complex layouts. Advanced models like Mistral OCR 4: Advanced Document Extraction and Multilingual Performance Summary Report leverage bounding box precision to handle 170+ languages and complex document structures, moving beyond basic text recognition to precise spatial localization of textual elements.