Deep Learning Object Detection is the computer vision task where neural networks identify and localize multiple objects within an image by predicting both class labels and bounding box coordinates — evolved from two-stage architectures (R-CNN family) that first propose regions then classify them, to one-stage detectors (YOLO, SSD) that predict directly in a single pass, and most recently to transformer-based detectors (DETR) that eliminate hand-crafted components like anchors and NMS.

Two-Stage Detectors

• R-CNN → Fast R-CNN → Faster R-CNN: The R-CNN lineage introduced region proposal networks (RPNs) that share convolutional features with the detection head. Faster R-CNN's RPN generates ~300 candidate regions per image; each region is classified and refined by a second-stage head. High accuracy but relatively slow (~5-15 FPS) due to the per-region computation.
• Cascade R-CNN: Multiple detection heads in series with progressively higher IoU thresholds, improving localization accuracy through iterative refinement.

One-Stage Detectors

• YOLO (You Only Look Once): Divides the image into a grid; each cell predicts bounding boxes and class probabilities in a single forward pass. YOLOv1 through YOLOv11 represent continuous evolution in backbone design, neck architecture (FPN, PANet), and training strategies. YOLOv8/v11 achieve >50 mAP on COCO at >100 FPS on GPU.
• SSD (Single Shot Detector): Predicts at multiple feature map scales, detecting small objects from high-resolution maps and large objects from low-resolution maps.
• Anchor-Free Detectors: FCOS, CenterNet predict object centers and distances to bounding box edges, eliminating anchor design (a major source of hyperparameter tuning). Most modern YOLO versions have adopted anchor-free prediction.

Transformer-Based Detection

• DETR (Detection Transformer): Uses a transformer encoder-decoder with learned object queries. Bipartite matching loss assigns predictions to ground truth without NMS. Eliminates anchors, NMS, and most hand-crafted components. Clean, end-to-end trainable.
• Deformable DETR: Adds deformable attention that attends to a sparse set of sampling points rather than all spatial locations, dramatically improving convergence speed (10x faster than DETR).
• RT-DETR: Real-time DETR variant that achieves YOLO-competitive speed by efficiently decoupling intra-scale and cross-scale feature interaction.

Backbone and Neck Architecture

• Feature Pyramid Network (FPN): Multi-scale feature maps with top-down pathway and lateral connections. Standard for detecting objects at different scales.
• Backbones: ResNet, CSPDarknet, EfficientNet, Swin Transformer — the feature extraction base that largely determines the speed-accuracy tradeoff.

Deep Learning Object Detection is the visual perception foundation that enables autonomous driving, robotic manipulation, medical imaging, and surveillance — having evolved from slow, multi-stage pipelines to real-time, end-to-end systems that detect hundreds of objects in a single image in milliseconds.