objektdeteksjon

Objektdeteksjon, the Norwegian term for object detection in computer vision, involves identifying objects of predefined classes in images or video and determining their positions. Outputs often include bounding boxes or segmentation masks, together with class labels and confidence scores for each detected instance.

Approaches are broadly categorized into two-stage detectors, which first generate region proposals and then classify them, and one-stage detectors, which predict objects and locations in a single pass. Examples: Faster R-CNN (two-stage), YOLO, SSD, RetinaNet (one-stage). Recent advances include transformer-based DETR.

Data representations include bounding boxes (x, y, width, height) or pixel-wise masks for segmentation. Evaluation commonly uses mean average precision (mAP) across Intersection-over-Union (IoU) thresholds. Popular benchmarks include COCO and PASCAL VOC.

Applications span autonomous driving, video surveillance, robotics, medical imaging, and quality control in manufacturing.

Challenges include occlusion, small object sizes, varying lighting, real-time constraints, dataset bias, and domain shift. Ongoing research seeks more accurate, efficient, and generalizable detectors, including lightweight models for edge devices and transformer-based architectures.