{
  "id": "yolo",
  "title": "YOLO Object Detection Models",
  "lang": "en",
  "source_lang": "en",
  "summary": "YOLO is a family of real-time object detection models that predict object locations and categories through a unified neural-network pipeline.",
  "category": "computer-science",
  "subcategory": "artificial-intelligence",
  "type": "product",
  "aliases": [
    "You Only Look Once",
    "YOLO detectors",
    "YOLO model family"
  ],
  "generated_at": "2026-11-02",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/yolo/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/yolo/",
      "json": "https://aiwiki.page/zh/yolo.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "YOLO, short for You Only Look Once, is a family of deep learning models for object detection: locating objects in an image and assigning them category labels. The original model, introduced by Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi in a 2015 preprint and presented at the 2016 Conference on Computer Vision and Pattern Recognition, formulated detection as a single prediction problem. Instead of first generating candidate regions and then classifying each region separately, it used one neural network to predict bounding boxes and class information directly from an image.\n\nYOLO became influential because it demonstrated a practical balance between detection accuracy and processing speed. Its successors introduced better feature extraction, multiscale prediction, training objectives, and deployment techniques, making YOLO-style systems widely used in video analytics, industrial inspection, robotics, and other areas of computer vision. The name now denotes a broad and partly branching ecosystem rather than one architecture maintained by a single research group. Different versions can differ substantially in their box representation, assignment rules, output heads, software licenses, and need for postprocessing."
    },
    {
      "heading": "Detection formulation and the original model",
      "text": "Object detection combines two tasks that are distinct in ordinary image classification: determining which categories are present and identifying where each instance appears. A detector commonly represents an object by an axis-aligned bounding box, a category, and a confidence score. For an image containing two cars and a pedestrian, the required output is three separate detections rather than a single image-level label.\n\nEarly successful deep detectors often followed a proposal-based pipeline. In R-CNN and its successors, candidate regions were generated or proposed and then classified and refined. This division could produce accurate detections but introduced additional processing stages. YOLO instead treated image-wide detection as a regression problem performed by one neural network. Its defining contribution was this unified formulation, not the literal elimination of all preprocessing or postprocessing.\n\nThe original model divided an image into an \\(S \\times S\\) grid. A grid cell was responsible for an object when the object’s center fell inside that cell. Each cell predicted \\(B\\) boxes, with five values per box: center coordinates \\(x,y\\), width \\(w\\), height \\(h\\), and a confidence value. It also predicted \\(C\\) conditional class probabilities shared by the cell’s boxes. The output therefore had the form\n\n\\[S \\times S \\times (5B+C).\\]\n\nFor the PASCAL VOC configuration, \\(S=7\\), \\(B=2\\), and \\(C=20\\), giving a \\(7 \\times 7 \\times 30\\) output. The network took a \\(448 \\times 448\\) image and used 24 convolutional layers followed by two fully connected layers. A smaller Fast YOLO model traded accuracy for speed.\n\nThe original confidence target combined object presence with intersection over union (IoU), a measure of overlap between a predicted box and its ground-truth box. At prediction time, class-conditional probabilities were multiplied by the box confidence to produce class-specific scores. This formulation differs from later implementations, so the meaning of a “YOLO confidence score” must be interpreted in relation to a particular model.\n\nOn PASCAL VOC 2007, the original paper reported 63.4% mean average precision at 45 frames per second for YOLO, and 52.7% at 155 frames per second for Fast YOLO, using its stated hardware and evaluation conditions. Its speed was accompanied by characteristic limitations: coarse spatial organization, difficulty with small or closely grouped objects, and localization errors. These weaknesses helped determine the design priorities of subsequent versions."
    },
    {
      "heading": "Historical development and branching lineages",
      "text": "YOLO’s history is not a single sequence in which each numbered release comes from the same authors. The first three major versions were associated with Redmon and collaborators. Later systems were developed by several research groups and companies, sometimes sharing code or architectural ideas and sometimes pursuing distinct research goals. A version number is therefore a name within an ecosystem, not a standardized indicator of compatibility or performance.\n\nYOLOv2 and YOLO9000, presented in 2017, replaced several features of the original design. YOLOv2 introduced an anchor-based prediction scheme, extensive use of batch normalization, a Darknet-19 backbone, and multiscale training. It used clustering of ground-truth box dimensions to obtain useful anchor shapes. YOLO9000 extended the work through joint training on detection and classification datasets, using a hierarchical category representation called WordTree. Its support for more than 9,000 categories did not mean that every category had equally extensive box-level supervision.\n\nYOLOv3, described by Redmon and Farhadi in 2018, introduced the Darknet-53 backbone and predictions at three spatial scales. It used independent logistic class predictions rather than requiring all class probabilities to sum to one. Its combination of residual feature extraction and multiscale outputs made it a durable baseline, especially for applications needing accessible training code and efficient inference.\n\nIn 2020, Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao introduced YOLOv4, combining architectural changes with an extensive set of training techniques. Its design included CSPDarknet53, spatial pyramid pooling, and path aggregation. Ultralytics separately released YOLOv5 in 2020 as a PyTorch-based software family. Its widespread use reflected not only detector design but also integrated training, validation, export, and deployment tools. YOLOv5 was not simply a paper-defined continuation by YOLO’s original authors.\n\nFurther branches included YOLOX in 2021, which emphasized an anchor-free formulation, decoupled prediction heads, and a sample-assignment method called SimOTA. YOLOv6, developed at Meituan, focused on industrial deployment. YOLOv7, introduced in 2022 by Wang, Bochkovskiy, and Liao, examined efficient network structures and training strategies. Ultralytics introduced YOLOv8 in 2023 with anchor-free detection and support for additional vision tasks.\n\nThe 2024 YOLOv9 work introduced programmable gradient information and the GELAN architecture. YOLOv10, also introduced in 2024, pursued efficient end-to-end detection without conventional non-maximum suppression through a dual-assignment training design. These examples illustrate different research directions rather than a universal migration path. Statements such as “YOLO uses anchors” or “YOLO requires suppression” can be correct for particular releases while being false for others."
    },
    {
      "heading": "Backbone, neck, and prediction heads",
      "text": "A useful way to describe many modern YOLO detectors is to divide them into a backbone, neck, and head. These are functional components rather than mandatory names for fixed blocks. The backbone extracts visual features; the neck combines information across resolutions; and the head converts those features into detection predictions. The original YOLO did not have the same explicit multiscale organization as later implementations.\n\nMost YOLO backbones are based primarily on convolutional neural networks. Early layers process relatively fine spatial patterns, while deeper layers combine them into more abstract features over larger receptive fields. Downsampling reduces the height and width of feature maps, lowering later computation but also discarding spatial detail. Residual and cross-stage connections help information and gradients move through deep networks without requiring every layer to reconstruct the complete representation independently.\n\nThe neck addresses the tension between semantic strength and spatial precision. Deep, low-resolution features are useful for recognizing what an object is; higher-resolution features preserve information about where small objects occur. Feature pyramid networks propagate semantic information toward finer resolutions, while path-aggregation designs add further routes for combining features. Spatial pyramid pooling aggregates context over different pooling windows. These mechanisms help a detector handle objects whose sizes differ substantially within one image.\n\nA common multiscale configuration predicts at strides of 8, 16, and 32 pixels relative to the input. For a \\(640 \\times 640\\) image, the corresponding feature maps are \\(80 \\times 80\\), \\(40 \\times 40\\), and \\(20 \\times 20\\). They contain 8,400 spatial locations in total. An anchor-based head using three anchors per location produces 25,200 candidate boxes before filtering; an anchor-free head may begin with one candidate per location. Additional high-resolution levels can improve small-object detection, but increase activation memory and computation.\n\nAnchor-based heads predict adjustments to predefined reference shapes. One YOLOv3-style parameterization is\n\n\\[b_x=\\sigma(t_x)+c_x,\\qquad b_y=\\sigma(t_y)+c_y,\\qquad b_w=p_w e^{t_w},\\qquad b_h=p_h e^{t_h},\\]\n\nwhere \\(c_x,c_y\\) identify a grid location, \\(p_w,p_h\\) are anchor dimensions in the corresponding coordinate system, and the \\(t\\) values are network outputs. These equations are not universal across YOLO versions.\n\nAnchor-free heads commonly predict distances from a feature-map location to the sides of a box, or use another representation without a fixed set of anchor shapes. Some predict distributions over distance bins rather than a single scalar for each side. Heads may also separate classification and localization into different branches. This decoupling recognizes that category recognition and precise boundary estimation do not necessarily benefit from identical features or optimization dynamics."
    },
    {
      "heading": "Training objectives and assignment of supervision",
      "text": "Training a YOLO detector normally uses supervised learning with images annotated by object categories and bounding boxes. The central difficulty is not just calculating an error for an existing prediction. The system must also decide which of thousands of candidate predictions should learn from each annotated object. This process, often called positive-sample assignment, strongly influences convergence, localization quality, and behavior in crowded scenes.\n\nIn the original YOLO, the grid cell containing an object’s center handled that object, and the predicted box with the highest overlap was selected as responsible for it. Anchor-based successors commonly assigned objects according to anchor shape and spatial position. More recent approaches can select positives dynamically using classification scores, localization quality, or a combined matching cost. One-to-many assignment gives several predictions supervision from the same object; one-to-one assignment trains a single matched prediction for each object. These alternatives create different trade-offs between learning efficiency and duplicate removal.\n\nA modern detection loss function can be represented schematically as\n\n\\[L=\\lambda_{\\mathrm{box}}L_{\\mathrm{box}} +\\lambda_{\\mathrm{cls}}L_{\\mathrm{cls}} +\\lambda_{\\mathrm{obj}}L_{\\mathrm{obj}} +\\lambda_{\\mathrm{aux}}L_{\\mathrm{aux}}.\\]\n\nThe coefficients weight box regression, classification, objectness, and any auxiliary terms. Not every implementation contains all four components. Some heads predict explicit objectness, while others integrate object presence or localization quality into their class scores.\n\nEarly YOLO used a sum-squared-error objective and applied square roots to box widths and heights to moderate the influence of large boxes. Later systems commonly use IoU-related localization losses. Generalized IoU, distance IoU, and complete IoU introduce information about enclosing regions, center distance, or aspect ratio to address weaknesses of overlap alone. Some distribution-based heads additionally use a distribution focal loss to supervise discretized box-edge distances. Classification losses are often based on binary cross-entropy, sometimes modified to account for imbalance or prediction quality.\n\nOptimization proceeds through backpropagation, usually with stochastic gradient descent or an adaptive optimizer. The learning-rate schedule, batch size, image resolution, and loss weights interact with architecture and assignment rules. Pretraining on ImageNet or another large dataset can supply useful visual features, while transfer learning adapts those features to a different category set or imaging domain.\n\nData augmentation is particularly prominent in YOLO training. Mosaic augmentation combines multiple images into a composite; other techniques alter scale, color, position, perspective, or image composition. These transformations expose the model to varied contexts and object sizes, but must preserve meaningful annotations. Excessive cropping can remove the evidence needed to identify an object, and inappropriate transformations can distort domain-specific cues. Annotation completeness also matters: a visible but unlabeled object may be treated as background, creating contradictory supervision rather than merely reducing the amount of available data."
    },
    {
      "heading": "Inference, confidence, and duplicate suppression",
      "text": "At inference time, a conventional YOLO pipeline preprocesses an image, runs the network, decodes the outputs, filters low-scoring candidates, and resolves duplicates. “Single-stage” refers principally to the detection architecture; it does not imply that an application consists of a single operation. Image decoding, resizing, memory movement, and postprocessing may contribute substantially to total latency.\n\nPreprocessing usually resizes the input to a supported resolution and converts pixel values into the numerical representation expected by the model. Letterboxing preserves the original aspect ratio by resizing the image and adding padding. The inverse transformation must be applied to predicted coordinates so that boxes align with the original image. Failure to account for padding or differences in scale can produce systematically misplaced boxes even when the neural network itself is functioning correctly.\n\nOutput decoding depends on the head. Anchor offsets must be transformed into box coordinates; distance distributions may be converted to expected distances; and class logits are passed through the appropriate activation function. An explicit objectness score may be multiplied by a class score, but this is not a universal rule. The resulting detection score is generally a ranking signal learned under a particular objective, not a guaranteed calibrated probability that a detection is correct.\n\nMost conventional YOLO implementations use non-maximum suppression (NMS). After sorting candidate detections by score, NMS retains a high-scoring box and removes lower-scoring boxes that overlap it too strongly. For boxes \\(A\\) and \\(B\\),\n\n\\[\\operatorname{IoU}(A,B)= \\frac{|A\\cap B|}{|A\\cup B|}.\\]\n\nSuppression is often performed separately for each class, although class-agnostic variants exist. A lower IoU threshold suppresses nearby boxes more aggressively; a higher threshold permits more overlap. In crowded scenes, aggressive suppression can remove valid detections of distinct objects. Conversely, permissive settings can leave duplicate predictions.\n\nThe confidence threshold creates a separate trade-off. Raising it generally removes more false positives but also more true objects. Lowering it increases the number of candidates passed to later processing and may improve recall at the expense of precision and runtime. Consequently, comparisons made at different thresholds need careful interpretation.\n\nNMS-free designs change the training and output structure so that one object is intended to produce one final prediction. YOLOv10, for example, uses one-to-many supervision to support learning alongside a one-to-one branch used for inference. Such designs avoid conventional overlap-based suppression, although they still require operations such as decoding and selecting outputs. “End-to-end” in this context describes the relation between prediction and matching more precisely than it describes the complete camera-to-application pipeline."
    },
    {
      "heading": "Benchmarks and the meaning of real-time performance",
      "text": "YOLO performance is usually described along at least two axes: detection accuracy and computational cost. Neither can be reduced to a version number. Input resolution, model size, dataset, training schedule, numerical precision, hardware, and software backend can all change the result. A small model from a newer family may be faster but less accurate than a large model from an older family.\n\nThe main accuracy measure is mean average precision (mAP). Predicted detections are matched to ground-truth objects using class agreement and an IoU threshold. Unmatched predictions count as false positives, while unmatched annotations contribute to false negatives. Varying the confidence threshold produces a precision–recall curve; average precision summarizes that curve for a class under the benchmark’s specified interpolation procedure. Mean average precision averages over classes and, in some protocols, over several overlap thresholds.\n\nPASCAL VOC evaluations commonly used an IoU threshold of 0.5. The COCO dataset established a more demanding detection metric that averages AP across IoU thresholds from 0.50 to 0.95 in steps of 0.05. COCO also reports measures for small, medium, and large objects. These distinctions matter: a model may identify the correct category and approximate location adequately at IoU 0.5 yet perform poorly when precise boundaries are required. VOC mAP and COCO AP therefore cannot be treated as interchangeable percentages.\n\nDataset handling is equally important. A validation set supports model selection and threshold tuning; a test set is intended for final evaluation. Repeatedly choosing settings based on test performance undermines the independence of that evaluation. In video datasets, randomly separating neighboring frames can cause data leakage, because training and evaluation may contain nearly identical scenes. A high score in that setting may not reflect performance on a new site, camera, or lighting condition.\n\nSpeed reporting requires an explicit measurement boundary. Network-only timing excludes some or all preprocessing, transfer, decoding, and suppression. End-to-end timing includes a larger portion of the application. Throughput in frames per second also differs from latency: batching can improve throughput while increasing the time an individual frame spends waiting.\n\nFor strictly serial processing with latency \\(T\\) seconds per frame, throughput is approximately \\(1/T\\), but pipelining and batching break this simple relationship. Hardware warm-up, synchronization, fixed versus variable shapes, and precision settings also affect measurements. Parameter count and multiply–accumulate estimates provide useful context, but neither alone predicts execution speed. A reproducible comparison therefore specifies the model checkpoint, input shape, hardware, runtime, batch size, precision, timing procedure, and whether postprocessing is included."
    },
    {
      "heading": "Deployment and computational optimization",
      "text": "YOLO deployment involves turning a trained model into a reliable component of a larger system. The training environment may use flexible operators, dynamic shapes, and debugging features that are unsuitable for a constrained runtime. Export and optimization translate the model into representations and kernels supported by the target platform, while preserving its expected numerical behavior.\n\nA common route exports a model to Open Neural Network Exchange (ONNX), then executes it with a compatible inference engine. On supported NVIDIA platforms, TensorRT can optimize execution through kernel selection, graph transformations, and reduced precision. Other platforms use different runtimes for desktop processors, mobile accelerators, or embedded devices. Export support is version-specific: a model can train successfully yet contain an operator that a particular backend cannot execute efficiently.\n\nReduced precision is one major optimization. FP16 inference can lower memory traffic and exploit suitable accelerator hardware. INT8 quantization represents many values with eight-bit integers, potentially reducing storage and execution cost further. Post-training quantization commonly uses representative calibration data to determine numerical ranges; quantization-aware training simulates relevant effects during optimization. Accuracy changes depend on the model, backend, and data. Small or low-confidence objects can be particularly vulnerable when quantization perturbs delicate feature responses.\n\nGraph transformations can improve efficiency without changing the intended computation. At inference, batch-normalization parameters can often be folded into a preceding convolution. Some architectures use training-time branches that can be algebraically reparameterized into a simpler inference structure. Such transformations distinguish the network used to learn parameters from the graph ultimately executed on the device.\n\nMemory and data movement can dominate practical performance. A graphics processing unit may execute the detector quickly while waiting on image decoding, host-to-device transfers, or a CPU-based postprocessor. On embedded hardware, activation tensors at high resolution can be more restrictive than parameter storage. A nominally lightweight model can perform poorly if its operations map inefficiently to the available accelerator.\n\nDeployment validation therefore concerns the whole pipeline. Exported outputs need to preserve class ordering, coordinate conventions, preprocessing, and decoding behavior. A system also needs to handle empty frames, unusual aspect ratios, variable scene density, and changing input rates. Video applications may prefer dropping outdated frames to accumulating an ever-growing queue, because a correct detection delivered too late can be operationally unhelpful. These engineering choices are separate from benchmark AP but directly affect the usefulness of the deployed detector."
    },
    {
      "heading": "Task extensions and practical applications",
      "text": "Although YOLO began as a bounding-box detector, its architectural approach has been extended to several related vision tasks. These extensions generally retain a shared feature extractor and add task-specific outputs. A package that offers detection, segmentation, and pose estimation under one name is therefore a family of trained models or heads, not necessarily one network that performs every task simultaneously.\n\nIn instance segmentation, each detected object receives a pixel-level mask. One efficient design predicts a set of image-wide mask prototypes and a small vector of coefficients for each instance; the coefficients combine the prototypes into an object-specific mask. This differs from semantic segmentation, which assigns categories to pixels without necessarily separating different instances of the same category. A detector-derived mask head must preserve enough spatial detail to distinguish boundaries, especially for thin structures or partially overlapping objects.\n\nPose-estimation variants add keypoint coordinates and sometimes visibility or confidence values. Human pose models commonly predict a predefined set of anatomical landmarks, whereas other datasets may define landmarks for animals, machinery, or manufactured parts. Oriented-box detection adds information about rotation, which is useful for elongated or arbitrarily oriented objects in aerial imagery, shipping scenes, and document analysis. Its overlap calculations and suppression rules differ from those for ordinary axis-aligned boxes.\n\nYOLO detections also serve as inputs to multi-object tracking. A tracker links detections across frames using motion predictions, spatial overlap, appearance information, or combinations of these signals. Detection alone does not provide persistent identity. If a person is missed for several frames, the tracker may terminate the track or assign a new identity when the person reappears. Thus, high per-frame AP does not guarantee accurate trajectories or counts over time.\n\nIndustrial uses include locating products on conveyors, identifying visible surface defects, and counting components. Traffic applications detect vehicles and pedestrians; agricultural applications locate fruit, plants, or animals. In these settings, domain-specific annotation often matters more than the general-purpose label set of a pretrained model. An object category that looks consistent in a benchmark may vary greatly with camera angle, weather, illumination, or production conditions.\n\nRobotic systems can combine detections with depth sensing, calibration, and motion planning. A two-dimensional box alone does not establish an object’s three-dimensional position, physical dimensions, or suitability for grasping. Likewise, a detection of a vehicle does not directly determine its speed. Those quantities require additional geometry or temporal measurements. YOLO’s practical role is often to supply a fast visual hypothesis to a larger system, whose other components establish identity, distance, timing, or action."
    },
    {
      "heading": "Limitations and continuing research",
      "text": "YOLO detectors remain sensitive to the information available in their input. A tiny object may occupy only a few pixels before downsampling and become indistinguishable from background in deeper feature maps. Occlusion removes visible evidence, while motion blur, glare, compression artifacts, and low illumination can alter the patterns learned during training. Higher input resolution and image tiling can recover some small-object performance, but increase computation and can create boundary effects or duplicate detections across tiles.\n\nCrowded scenes expose interactions between representation, assignment, and postprocessing. Multiple objects may compete for nearby prediction locations, and heavy overlap can make a correct detection resemble a duplicate. Anchor-free prediction removes the need to select predefined anchor shapes, but does not eliminate ambiguity about which spatial locations should represent which objects. Similarly, removing NMS transfers more responsibility to training and matching rather than making duplicate prediction an irrelevant problem.\n\nDistribution shift is another central limitation. A detector trained on daylight road scenes may behave differently at night or in another geographic setting. Changes in camera optics, viewpoint, image processing, or category definitions can also affect performance. This is distinct from ordinary overfitting to training examples: a model may generalize well within one benchmark distribution and still fail after deployment because the operating conditions differ.\n\nMost conventional YOLO models use a closed category vocabulary. An unfamiliar object may be ignored or assigned to a visually similar known class. A low score is not a reliable universal indicator of novelty, and a high score does not prove that the object belongs to the training taxonomy. Open-vocabulary variants seek to connect region features to language-derived category representations, bringing detection into contact with multimodal learning. Their behavior depends on the language–vision model, prompts, training categories, and evaluation protocol as well as on the detector backbone.\n\nSecurity and privacy concerns arise from the surrounding application and the model’s technical limitations. Deliberately constructed visual patterns can interfere with learned detectors under some conditions, although effectiveness depends on viewpoint, printing, illumination, and the system being tested. Video analytics can also process identifiable people or sensitive locations even when the detector outputs only boxes and category labels. Detection performance does not by itself establish the appropriateness of collecting or retaining the underlying imagery.\n\nActive research investigates better supervision assignment, more efficient multiscale features, improved confidence calibration, temporal information, open-vocabulary recognition, and hardware-aware architecture design. Transformer-based detectors provide an important alternative, particularly through set prediction and learned matching, while some YOLO variants incorporate attention components. The resulting boundary is increasingly architectural rather than categorical: real-time detection research compares different ways to allocate computation, associate predictions with objects, and preserve localization accuracy under limited latency and memory."
    }
  ],
  "links": [
    "deep-learning",
    "object-detection",
    "computer-vision",
    "image-classification",
    "artificial-neural-network",
    "intersection-over-union",
    "batch-normalization",
    "pytorch",
    "convolutional-neural-network",
    "supervised-learning",
    "loss-function",
    "backpropagation",
    "learning-rate",
    "imagenet",
    "transfer-learning",
    "data-augmentation",
    "activation-function",
    "non-maximum-suppression",
    "coco-dataset",
    "validation-set",
    "test-set",
    "data-leakage",
    "open-neural-network-exchange",
    "graphics-processing-unit",
    "instance-segmentation",
    "semantic-segmentation",
    "distribution-shift",
    "overfitting",
    "multimodal-learning"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}