object detection

**Object Detection** is the **computer vision task of localizing and classifying all objects in an image** — outputting bounding boxes and class labels, and serving as the foundation for autonomous driving, surveillance, robotics, and medical imaging. **Detection Paradigms** **Two-Stage (R-CNN Family)**: - Stage 1: Region Proposal Network (RPN) → generate candidate regions. - Stage 2: Classify and refine each region independently. - Examples: R-CNN, Fast R-CNN, Faster R-CNN, Mask R-CNN. - Pros: Higher accuracy. Cons: Slower (~5 FPS). **One-Stage (YOLO Family)**: - Single forward pass predicts all boxes simultaneously. - Divide image into SxS grid; each cell predicts B bounding boxes. - YOLOv1 (2016) → YOLOv8 (2023): Accuracy improved to match two-stage. - YOLOv8: 50+ FPS on GPU, 55 mAP on COCO — standard for real-time detection. **Anchor-Based vs. Anchor-Free** - **Anchor boxes**: Predefined aspect ratios/sizes. Network predicts offsets from anchors. - Problem: Anchor hyperparameters, many candidates, slow. - **Anchor-free (FCOS, CenterNet)**: Predict from center or feature point directly. - Simpler, faster, better on objects with unusual aspect ratios. **Feature Pyramid Network (FPN)** - Multi-scale feature extraction: Top-down pathway with lateral connections. - Small objects detected at high-resolution features (early layers). - Large objects detected at low-resolution features (later layers). - Standard in all modern detectors. **DETR (Detection Transformer, 2020)** - Transformer encoder-decoder with learned object queries. - No anchors, no NMS — set prediction with Hungarian matching loss. - Global attention captures long-range relationships. - Deformable DETR: 10x faster convergence with deformable attention. **Key Metrics** - **mAP (mean Average Precision)**: Standard benchmark metric at IoU thresholds. - COCO dataset: mAP@[.5:.95] — standard benchmark. - State-of-art (2024): 60+ mAP with ensemble/large models. Object detection is **the gateway task for visual understanding of scenes** — its algorithms power every camera-based safety system, content moderation tool, and autonomous navigation system deployed at scale today.