Object Detection

A detector predicts class scores and spatial regions, usually bounding boxes, for a variable number of objects. Two-stage systems propose regions and classify or refine them; one-stage systems predict detections more directly from image features. Postprocessing filters low-confidence candidates and may suppress overlapping duplicate predictions. Training needs instance annotations whose coordinates follow a precise convention. Evaluation matches predictions with reference objects using overlap criteria and then measures precision and recall across confidence thresholds. Object detection differs from segmentation, which assigns detailed pixels, and from tracking, which associates instances over time. Its output supports counting or localization only within the defined categories and annotation rules.

Write class and box-boundary guidelines, including partially visible objects and difficult examples. Audit annotation completeness and convert coordinates carefully after resizing or cropping. Split data by scene, location or video sequence, not adjacent frames. Compare detector quality by object size, occlusion and class, using average precision plus application-specific false-positive and missed-object costs. Set thresholds on development data and measure latency with postprocessing included. Inspect overlays to distinguish localization errors from label errors. The deliverable is an evaluated detector and documented operating threshold whose spatial outputs match the downstream coordinate system.

Detection confidence is not automatically a calibrated probability, and overlap metrics may not reflect downstream counting errors. Unannotated objects can make correct predictions appear false, while incomplete labels corrupt training. Small, occluded or unfamiliar objects are common failures. Suppression can remove legitimate nearby instances. A successful detector does not establish object identity over time or detailed shape. Evaluate class coverage, spatial accuracy and threshold behavior separately, including scenes that contain no target objects.