What is Object Detection?
Object detection locates instances of defined classes within an image or video frame. A detector usually returns a class label, confidence score, and bounding box for each candidate.
How Object Detection works
An object detector processes visual input and proposes spatial regions associated with classes from its training taxonomy. Post-processing filters low scores and usually suppresses overlapping candidates that appear to describe the same instance. Results can feed trackers, redaction, smart crops, inventory counts, or human annotation rather than serving as final truth. Performance depends on object scale, scene domain, annotation policy, and operating threshold, so validation must reflect the actual media stream.
Key facts
- Non-maximum suppression can remove duplicate overlapping boxes, but aggressive overlap settings may also discard distinct objects standing close together.
- Bounding boxes localize rectangular extents rather than exact silhouettes; applications needing pixel boundaries require segmentation or another refinement stage.
- Mean average precision summarizes ranked detections across classes and overlap thresholds, but it does not directly encode an application’s cost of each error.
When Object Detection matters
Run detection before cropping, moderation, counting, or analysis when the location of subjects matters. Confidence thresholds trade missed objects against false detections and should reflect downstream risk.