What is object detection
Object detection needs to answer two questions: what objects are in the picture (classification), and where they are located (localization). Localization is usually represented by a rectangular bounding box. It is one of the most fundamental and important tasks in computer vision.
Two-stage approach
Starting from R-CNN (Girshick et al., 2014), this type of method first generates a batch of candidate regions, and then classifies and regresses the bounding boxes for each region. Fast R-CNN (2015) and Faster R-CNN (2015) are continuously optimized, with the latter using Region Proposal Networks (RPNs) to hand over candidate generation to neural networks, significantly speeding up the process.
Single-stage method
YOLO (Redmon et al., 2015) treats detection as a one-time regression problem, directly predicting object categories and positions on the grid with extremely fast speed; SSD (Liu et al., 2016) performs detection on multi-scale feature maps, balancing speed and accuracy. The single-stage method has promoted the popularization of detection in real-time scenarios.
Several key concepts
IoU (Intersection over Union) measures the degree of overlap between the predicted box and the true box; Non maximum suppression (NMS) is used to remove duplicate boxes on the same object; MAP (Mean Average Precision) is the most commonly used evaluation metric for detection tasks. Early methods still commonly relied on manually designed anchors.
Modern progress
DETR (Carion et al., 2020) used Transformer to transform detection into end-to-end ensemble prediction, removing anchor boxes and NMS. Subsequent work has continuously improved its convergence speed and accuracy. In addition, stronger backbone networks such as FPN multi-scale feature pyramids and self supervised pre training continue to improve detection performance.
Typical Applications
Autonomous driving perception, security monitoring, industrial quality inspection, medical image analysis, shelf recognition for retail and warehousing, etc.
Reference source
Comprehensive compilation of publicly available academic literature, including Girshick et al. (2014, R-CNN), Ren et al. (2015, Faster R-CNN), Redmon et al. (2015, YOLO), Liu et al. (2016, SSD), Carion et al. (2020, DETR), etc.