What is Bounding Box

Formats

• Corners: (x_min, y_min, x_max, y_max), used by Pascal VOC and most detection libraries
• COCO: (x, y, width, height) from the top-left corner
• YOLO: (x_center, y_center, width, height), each divided by the image size so values run from 0 to 1
• Rotated boxes add an angle, used for text lines and aerial images
• Converting between formats is the most common source of boxes that land in the wrong place

Intersection over Union (IoU)

IoU measures how well a predicted box matches a labeled one: the area where they overlap divided by the area they cover together. A predicted box of 100 × 100 pixels shifted 20 pixels right of an identical labeled box overlaps 80 × 100 = 8,000 pixels out of a union of 12,000, so IoU = 0.67. Detection benchmarks count a prediction as correct above a threshold, commonly 0.5; see mean average precision.

Bounding Boxes in Mixpeek

• The face identity extractor stores a `bbox` (x1, y1, x2, y2, width, height) and a detection score with every face it finds in images and video
• The document extractor stores a `bbox` for every layout block of a PDF page (paragraph, table, figure, form) and can render a thumbnail cropped to it
• Search results carry these boxes, so an application can highlight the matched face or passage instead of showing the whole frame or page

Common Pitfalls

• Mixing coordinate formats or origins between a model and the code that draws the boxes
• Forgetting to rescale boxes after resizing or cropping an image
• Using axis-aligned boxes for long rotated objects, which then cover mostly background
• Reporting detection accuracy without the IoU threshold it was measured at