You Only Look Once: Unified, Real-Time Object Detection (YOLO)

Detects and localizes multiple objects in images by dividing the input into an SxS grid and predicting bounding boxes and class probabilities directly from the full image in one forward pass. Uses a unified CNN architecture that jointly optimizes localization (bounding box coordinates) and classification (object class) end-to-end, eliminating the multi-stage pipeline of prior detectors. The regression-based approach treats detection as a direct coordinate prediction problem rather than region proposal refinement.

Extracts hierarchical spatial features from input images using a deep CNN backbone (typically 24 convolutional layers followed by 2 fully-connected layers) that progressively reduces spatial dimensions while increasing feature depth. Features at multiple scales implicitly capture both fine-grained details (early layers) and semantic context (deep layers), enabling detection of objects across a range of sizes. The architecture uses 1x1 convolutions for dimensionality reduction and 3x3 convolutions for spatial feature learning.

Simultaneously predicts bounding box coordinates (x, y, width, height) and class probabilities for each grid cell using a unified loss function that combines L2 regression loss for localization with cross-entropy classification loss. The loss function applies different weighting to localization and classification errors, with higher weight on localization errors in cells containing objects and classification errors in cells with objects. This joint optimization forces the network to learn both tasks end-to-end without separate training stages.

Executes complete object detection (feature extraction + localization + classification) in a single forward pass through a relatively shallow CNN (24 conv layers vs 50+ in ResNet), achieving 45-155 FPS on NVIDIA GPUs depending on model variant. The architecture avoids expensive operations like region proposal generation (RPN) and non-maximum suppression (NMS) post-processing, enabling inference latency <30ms on commodity hardware. Inference can be further accelerated through quantization, pruning, or deployment on mobile/edge devices.

Divides input images into an SxS grid (typically 7x7 for 448x448 input) and predicts bounding boxes directly from each grid cell without explicit anchor boxes. Each cell predicts B bounding boxes (typically 2) with coordinates (x, y, w, h) normalized relative to the cell, plus confidence scores and class probabilities. The grid-based approach implicitly anchors predictions to cell centers, enabling spatial awareness without explicit anchor generation. Bounding boxes can extend beyond cell boundaries, allowing detection of objects spanning multiple cells.

Removes redundant overlapping bounding box predictions after inference using intersection-over-union (IoU) thresholding. The algorithm sorts predictions by confidence score, greedily selects highest-confidence boxes, and suppresses lower-confidence boxes with IoU > threshold (typically 0.5) relative to selected boxes. This post-processing step is applied after decoding grid predictions to final image coordinates, reducing false positives from multiple overlapping detections of the same object.