Object Detection in 20 Years: A Survey

Zhengxia Zou; Keyan Chen; Zhenwei Shi; Yuhong Guo; Jieping Ye

Source-linked AI summary

Object Detection in 20 Years: A Survey

Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, Jieping Ye

arXiv:1905.05055v3May 13, 2019cs.CV

TL;DR

Object detection faces varied challenges, including rotation, scale changes, localization, occlusion, and speed. This survey maps the field’s technical evolution over 20 years, reviewing milestone detectors, technologies, datasets, metrics, acceleration methods, and future directions.

Problem
Object detection must address varied challenges, including rotation, scale changes, accurate localization, dense or occluded objects, and detection speed.
Method
The survey synthesizes object detection’s technical evolution alongside key technologies, state-of-the-art methods, datasets, metrics, and speed-up techniques.
Results
The paper reviews milestone detectors, key technologies, speed-up methods, datasets, and metrics across object detection’s 20-year history.
Takeaways & Limitations
The survey provides a roadmap for understanding object detection’s evolution and identifying future directions in this fast-moving field.

Abstract

from arXiv · show

Object detection, as of one the most fundamental and challenging problems in computer vision, has received great attention in recent years. Over the past two decades, we have seen a rapid technological evolution of object detection and its profound impact on the entire computer vision field. If we consider today's object detection technique as a revolution driven by deep learning, then back in the 1990s, we would see the ingenious thinking and long-term perspective design of early computer vision. This paper extensively reviews this fast-moving research field in the light of technical evolution, spanning over a quarter-century's time (from the 1990s to 2022). A number of topics have been covered in this paper, including the milestone detectors in history, detection datasets, metrics, fundamental building blocks of the detection system, speed-up techniques, and the recent state-of-the-art detection methods.

I. INTRODUCTION · II. OBJECT DETECTION IN 20 YEARS · A. A Road Map of Object Detection

Object detection identifies what objects are where and supports many computer vision tasks and real-world applications. This survey traces its evolution across traditional and deep-learning periods while reviewing milestones, datasets, metrics, key technologies, and speed-up techniques.

I. INTRODUCTION: Object detection identifies instances of visual objects by class and location, providing basic knowledge for computer vision applications.The passage frames the task around determining what objects are where in digital images.
I. INTRODUCTION: Object detection underpins instance segmentation, image captioning, and object tracking, while deep learning has driven major progress and increased research attention.The passage also notes applications including autonomous driving and robotics.
I. INTRODUCTION: Detection must address viewpoint, illumination, intraclass variation, rotation, scale changes, localization, dense and occluded objects, and speed constraints.These challenges differ across detection tasks because their objectives and constraints vary.
I. INTRODUCTION: The survey emphasizes technical evolution, key technologies, recent state-of-the-art methods, and comprehensive analysis of detection speed-up techniques.Its organizing perspective connects the past, present, and future of object detection.
II. OBJECT DETECTION IN 20 YEARS: The 20-year review covers milestone detectors, datasets, metrics, and the evolution of key detection techniques.These topics are presented as multiple views of object detection history.
II. OBJECT DETECTION IN 20 YEARS: The roadmap lists milestone detectors spanning VJ Det., HOG Det., DPM, R-CNN, SPPNet, Fast R-CNN, Faster R-CNN, YOLO, SSD, FPN, Retina-Net, CornerNet, CenterNet, and DETR.The figure identifies these detectors with their associated references.
A. A Road Map of Object Detection: Object detection progress is commonly divided into a traditional period before 2014 and a deep-learning-based period after 2014.The survey uses milestone emergence time and performance to highlight the technologies driving this historical transition.

1) Milestones: Traditional Detectors: · 2) Milestones: CNN based Two-stage Detectors:

Traditional object detection progressed from handcrafted features and sliding-window methods to influential models such as HOG and DPM, whose design principles shaped later detectors. After handcrafted features plateaued, CNN-based two-stage detectors introduced a coarse-to-fine pipeline and rapidly improved speed, accuracy, and multiscale representation.

1) Milestones: Traditional Detectors:: Early object detectors relied on handcrafted features and sophisticated speed-up techniques because effective image representations were unavailable.
1) Milestones: Traditional Detectors:: In 2001, the Viola–Jones detector achieved unconstrained real-time face detection and was tens or hundreds of times faster than contemporaneous methods at comparable accuracy.It used sliding windows to examine image locations and scales.
1) Milestones: Traditional Detectors:: HOG, introduced in 2005, balanced feature invariance and nonlinearity through dense-cell computation and overlapping block-based contrast normalization.Its primary motivation was pedestrian detection, and multiscale detection used repeated image rescaling.
1) Milestones: Traditional Detectors:: DPM, proposed in 2008 and winner of the VOC-07, -08, and -09 challenges, exemplified traditional detection through a divide-and-conquer decomposition of objects.Later detectors retained insights including mixture models, hard negative mining, bounding box regression, and context priming.
2) Milestones: CNN based Two-stage Detectors:: After handcrafted features saturated and detection research plateaued after 2010, CNNs revived the field by learning robust, high-level image representations.RCNN’s 2014 introduction marked the rapid evolution of deep-learning detection.
2) Milestones: CNN based Two-stage Detectors:: Two-stage detectors frame detection as coarse-to-fine: RCNN generates selective-search proposals, extracts CNN features for resized regions, and classifies them with linear SVMs.
2) Milestones: CNN based Two-stage Detectors:: SPPNet computed convolutional feature maps once for the entire image and generated fixed-length representations for arbitrary regions, becoming more than 20 times faster than R-CNN without sacrificing accuracy.On VOC07, SPPNet achieved mAP=59.2%.
2) Milestones: CNN based Two-stage Detectors:: Fast RCNN jointly trained detection and bounding-box regression, raising VOC07 mAP from 58.5% to 70.0% and running over 200 times faster than R-CNN.Faster RCNN then introduced an RPN for nearly cost-free proposals and achieved near-real-time performance at 17fps with ZF-Net.

3) Milestones: CNN based One-stage Detectors: · B. Object Detection Datasets and Metrics · 1) Datasets:

The paper contrasts fast, deployment-friendly one-stage detectors with accurate but complex two-stage methods, tracing milestones from YOLO through DETR. It also reviews major detection datasets, whose scale, categories, and annotations support progress in detection algorithms.

3) Milestones: CNN based One-stage Detectors:: Two-stage detectors use coarse-to-fine processing for recall and localization, whereas one-stage detectors offer one-step inference but struggle with dense and small objects.Two-stage methods achieve high precision but are hindered by poor speed and complexity; one-stage methods are favored for real-time mobile deployment.
3) Milestones: CNN based One-stage Detectors:: YOLO introduced deep-learning-era one-stage detection in 2015 by applying one neural network to the full image, reaching 155fps at VOC07 mAP=52.7%.Its enhanced version reached 45fps with VOC07 mAP=63.4%.
3) Milestones: CNN based One-stage Detectors:: SSD improved one-stage accuracy, especially for small objects, through multi-reference and multi-resolution detection, achieving COCO mAP@.5=46.5% at 59fps in its fast version.SSD was proposed in 2015 and combined detection speed and accuracy advantages.
3) Milestones: CNN based One-stage Detectors:: RetinaNet attributed the longstanding accuracy gap to extreme foreground-background imbalance and introduced focal loss by reshaping standard cross entropy.The method targeted the central training difficulty of dense one-stage detectors.
3) Milestones: CNN based One-stage Detectors:: CornerNet replaced anchor-box detection with keypoint prediction and corner re-grouping, outperforming most one-stage detectors of its time with COCO mAP@.5=57.8%.This approach avoided the category imbalance, hand-designed hyperparameters, and convergence burden associated with many reference boxes.
3) Milestones: CNN based One-stage Detectors:: CenterNet represented each object as a center point, regressed its attributes from that point, and removed keypoint grouping and NMS for fully end-to-end detection.It followed a keypoint-based paradigm while eliminating costly post-processing.
3) Milestones: CNN based One-stage Detectors:: DETR framed object detection as set prediction and proposed an end-to-end Transformer-based detection network in 2020.Transformers replace traditional convolution with attention-alone calculation to obtain a global-scale receptive field.
1) Datasets:: Advanced detection algorithms depend on larger, less-biased datasets, including PASCAL VOC, ILSVRC, MS-COCO, Open Images, and Objects365.These benchmarks provide varied object categories, instances, images, bounding boxes, and—in some cases—additional annotations or visual-relationship tasks.

2) Metrics: · C. Technical Evolution in Object Detection · 1) Technical Evolution of Multi-Scale Detection:

Object detection evaluation evolved from early task-specific measures to AP/mAP and COCO’s multi-threshold AP, while detection systems progressed through distinct multi-scale strategies. These strategies moved from feature pyramids and proposals toward deep regression, keypoints, anchors, and multi-resolution designs, alongside increasingly explicit use of context.

2) Metrics:: Early pedestrian detection commonly used miss rate versus false positives per window, but per-window measurement could fail to predict full-image performance.The Caltech pedestrian dataset later introduced a more comprehensive evaluation framework, although the supplied passage is truncated before its details.
2) Metrics:: AP, introduced in VOC2007, averages detection precision across different recall levels, while mean AP averages AP over categories.IoU between predicted and ground-truth boxes determines whether detections exceed a predefined threshold; 0.5-IoU mAP became the de facto metric.
2) Metrics:: COCO AP averages performance across IoU thresholds from 0.5 to 0.95, emphasizing more accurate object localization than a fixed threshold.This localization emphasis is relevant to applications such as robotic grasping.
1) Technical Evolution of Multi-Scale Detection:: Multi-scale detection addresses objects with different sizes and aspect ratios and passed through multiple historical periods over the past 20 years.The survey presents this evolution as one of the main technical challenges in object detection.
1) Technical Evolution of Multi-Scale Detection:: From 2004, feature-pyramid and sliding-window detectors such as HOG and DPM provided a major detection paradigm but paid little attention to different aspect ratios.These systems repeatedly scanned images with a fixed-size detection window.
1) Technical Evolution of Multi-Scale Detection:: Object proposals reduced exhaustive image-wide sliding-window search by supplying class-agnostic reference boxes likely to contain objects.Early proposal methods followed a bottom-up philosophy, while deep CNN recognition after 2014 supported a subsequent shift toward top-down detection.
1) Technical Evolution of Multi-Scale Detection:: Deep regression directly predicts bounding-box coordinates from deep-learning features, and post-2018 keypoint methods detect corners, centers, or representative points before objectwise grouping.The supplied passage identifies these as approaches that made multi-scale detection more straightforward with increased GPU computing power.
1) Technical Evolution of Multi-Scale Detection:: Multi-reference detection defines anchors or points at each image location, while multi-resolution detection uses different resolutions to detect objects.The survey describes multi-reference detection as the most used current method for multi-scale detection and lists both approaches among contemporary techniques.

2) Technical Evolution of Context Priming:

Context priming in object detection evolved from bootstrap-based hard negative mining to methods addressing severe background–object imbalance. Early detectors iteratively added misclassified backgrounds, while deep-learning detectors first balanced positive and negative weights before recognizing that this was insufficient.

Imbalance and hard negative mining: Object-detector training can face background–object imbalance as extreme as 10^7:1, allowing easy negatives to overwhelm learning.Hard negative mining was developed to address this problem.
Bootstrap: Bootstrap training began with a small background subset and iteratively added misclassified samples to reduce computation and address data imbalance.It was used in early detectors and later became standard in DPM and HOG detectors.
Deep-learning detectors: During 2014–2016, deep-learning detectors discarded bootstrap; Faster RCNN and YOLO instead balanced positive and negative window weights, but this did not completely solve imbalance.Researchers subsequently reintroduced bootstrap to object detection.

3) Technical Evolution of Hard Negative Mining: · 4) Technical Evolution of Loss Function:

The loss function supervises object detection through classification and localization losses, evolving from early L2 formulations toward losses that better model distributions, outliers, box geometry, and overlap.

4) Technical Evolution of Loss Function:: Classification and localization losses jointly supervise object detection by measuring prediction errors and enabling gradient-based weight updates.The loss is computed from deviations between predictions and true labels, after which backpropagation updates model weights.
4) Technical Evolution of Loss Function:: Early detectors such as YOLOv1 and YOLOv2 used MSE/L2 classification loss, while later methods typically adopted CE loss.L2 measures Euclidean-space error, whereas CE measures distribution differences as a form of likelihood.
4) Technical Evolution of Loss Function:: Localization loss evolved from prevalent L2 loss, which is sensitive to outliers and gradient explosion, to Smooth L1 loss combining L1 and L2 benefits.Smooth L1 was proposed to address limitations of the earlier localization formulation.
4) Technical Evolution of Loss Function:: IoU loss was introduced because treating bounding-box coordinates independently ignores their correlation and equal Smooth L1 errors can produce different IoU values.The four coordinates are x, y, w, and h, while IoU measures correspondence between prediction and ground truth.
4) Technical Evolution of Loss Function:: G-IoU improved IoU loss when non-overlapping boxes yielded IoU = 0 and therefore could not be optimized.This specifically addresses the non-overlapping-box case.
4) Technical Evolution of Loss Function:: DIoU extends IoU- and G-IoU-based regression by incorporating center-point distance alongside overlap area and aspect ratio.Distance-IoU identifies overlap area, center point distance, and aspect ratio as three geometric metrics for successful detection regression loss.

5) Technical Evolution of Non-Maximum Suppression: … D. Lightweight Network Design

The paper traces NMS from greedy post-processing to aggregation, learned, and NMS-free approaches, while organizing detector acceleration around the pipeline, backbone, and numerical computation. It highlights shared feature computation, cascades, pruning or quantization, and lightweight network design as speed-up strategies.

5) Technical Evolution of Non-Maximum Suppression:: NMS evolved from greedy score-based suppression of overlapping boxes toward methods that address imperfect top-scoring selections and replicated detections.Greedy NMS selects the highest-scoring box and removes neighbors above a predefined overlap threshold; early detectors did not always integrate NMS because the desired output was unclear.
5) Technical Evolution of Non-Maximum Suppression:: Bounding-box aggregation combines or clusters overlapping boxes into one detection while considering object relationships and spatial layout.The approach was used by the VJ detector and Overfeat, which won the ILSVRC-13 localization task.
5) Technical Evolution of Non-Maximum Suppression:: Learning-based NMS re-scores raw detections through an end-to-end network or a model trained to imitate NMS, improving occlusion and dense-object detection.These methods treat NMS as a learned filter rather than a traditional hand-crafted procedure.
5) Technical Evolution of Non-Maximum Suppression:: NMS-free detectors use one-to-one label assignment so each object has one prediction box, commonly training with the highest-quality box.This design enables fully end-to-end object-detection training and is described as more similar to human visual perception.
III. SPEED-UP OF DETECTION: Detector acceleration is grouped into speed-up of the detection pipeline, detector backbone, and numerical computation, with shared feature computation reducing repeated extraction.Computing the whole-image feature map only once has achieved tens or even hundreds of times of acceleration.
B. Cascaded Detection: Cascaded detection applies coarse-to-fine processing by filtering simple background windows cheaply before analyzing difficult windows with complex calculations.It is especially used for small objects in large scenes, including face and pedestrian detection.
D. Lightweight Network Design: Lightweight CNN detector design uses principles such as fewer channels and more layers, alongside additional methods proposed in recent years.These methods directly design compact networks rather than modifying an existing model after training.

1) Factorizing Convolutions: … 1) Speed Up with Integral Image:

The surveyed speed-up techniques reduce object-detection computation through convolution factorization, channel grouping, bottleneck compression, neural architecture search, numerical acceleration, and integral-image-based feature computation. These methods include theoretical complexity reductions and, for integral HOG maps, dozens of times’ acceleration without accuracy loss.

1) Factorizing Convolutions:: Factorizing convolutions builds lightweight CNNs by replacing large filters with multiple smaller filters or factorizing the channel dimension.A 7x7 filter can be replaced by three 3x3 filters with the same receptive field and greater efficiency.
2) Group Convolution:: Group convolution divides feature channels into m groups and processes them independently, theoretically reducing computation to 1/m of the original.The parameter count is also reduced by partitioning the feature channels.
3) Depth-wise Separable Convolution:: Depth-wise separable convolution is group convolution with groups equal to channels, reducing computation from O(dk2c) to O(ck2)+O(dc).1x1 filters typically transform the output to the desired number of channels.
4) Bottle-neck Design:: Bottleneck designs compress detector inputs or feature maps, reducing computation at the beginning of detection or in subsequent processing.The design uses layers containing few nodes compared with preceding layers.
5) Detection with NAS:: Neural architecture search reduces human intervention in designing detector backbones and anchor-based components by searching candidate networks efficiently and validating results at low cost.NAS addresses the candidate-network space, search strategies, and validation cost.
E. Numerical Acceleration: Numerical acceleration targets faster object detectors from the lowest implementation level.It is presented as a separate speed-up category in the survey.
1) Speed Up with Integral Image:: Integral images rapidly calculate sums over image subregions by exploiting integral-differential separability of convolution.Acceleration follows when dg(x)/dx is a sparse signal, enabling convolution through the right-hand part of the separable formulation.
1) Speed Up with Integral Image:: Dozens of times’ acceleration without losing any accuracy was achieved for pedestrian detection by computing integral HOG maps.Integral images also accelerate color and gradient histograms and allow histogram features at any location and size with constant computational complexity.

2) Speed Up in Frequency Domain: · IV. RECENT ADVANCES IN OBJECT DETECTION

The paper describes frequency-domain methods for accelerating convolutional operations in object detection, then transitions to reviewing recent state-of-the-art detection algorithms while noting that core principles remain stable.

2) Speed Up in Frequency Domain:: Convolution implements the window-wise inner product between a feature map and detector weights in linear object detection.This connects detector computation to convolutional operations.
2) Speed Up in Frequency Domain:: The Fourier transform provides a practical route to speeding up convolutions through the convolution theorem.The method is grounded in signal-processing properties of convolution and Fourier transformation.
2) Speed Up in Frequency Domain:: The frequency-domain calculation uses the Fourier transform, inverse Fourier transform, and point-wise product.These operations are represented by F, F −1, and ⊙, respectively.
2) Speed Up in Frequency Domain:: Fast Fourier Transform and Inverse FFT accelerate the frequency-domain convolution calculation.FFT and IFFT provide computational acceleration for the corresponding transforms.
2) Speed Up in Frequency Domain:: Vector Quantization approximates large data distributions with a small set of prototype vectors for compression and faster inner products.In object detection, VQ can accelerate the inner-product operation while also supporting data compression.
IV. RECENT ADVANCES IN OBJECT DETECTION: Recent advances focus on state-of-the-art object-detection algorithms over a short time range, following a review of two decades of technological evolution.The paper states that new technologies substantially influenced detection while fundamental principles and underlying logic remained unchanged.

3) Vector Quantization: … D. Improvements of Localization

The section surveys localization-oriented detection developments, including key-point formulations, robustness to rotation and scale changes, backbone effects, and localization improvements. It also contrasts single-resolution and multi-resolution training strategies for multi-scale detection.

3) Vector Quantization:: Multi-scale detection can train on a single resolution while backpropagating all object scales, or use image pyramids while selecting which object-scale gradients to retain.In multi-resolution training, gradients for objects that are too large or too small are discarded.
A. Beyond Sliding Window Detection: Detection can be framed as pair-wise localization of an object’s upper-left and lower-right ground-truth corners, with later methods adding center, extreme, or representative points.One implementation predicts corner heat maps.
B. Robust Detection of Rotation and Scale Changes: Recent research addresses robust detection under rotation and scale changes.The paper treats rotation and scale robustness as distinct directions of recent work.
1) Rotation Robust Detection:: Rotation robustness uses data augmentation, orientation-specific detectors, or rotation-invariant loss functions.These approaches target common rotation cases including face, text, and remote-sensing object detection.
2) Scale Robust Detection:: Scale-adaptive training addresses scale imbalance by constructing image pyramids during training and detection and selectively backpropagating object losses.Fixed-size image resizing with losses from all object scales can create scale imbalance, while detection-time pyramids alleviate rather than fundamentally solve it.
2) Scale Robust Detection:: Scale-adaptive detection compensates for unexpected scale changes through adaptive zoom-in methods that enlarge small objects and predictions of object scale distributions.Anchor size and aspect-ratio configurations are otherwise difficult to adapt to unexpected scale changes.
C. Detection with Better Backbones: Detection accuracy and speed depend heavily on feature-extraction backbones, including ResNet, CSPNet, Hourglass, and Swin Transformer.The cited systems include Faster RCNN, RFCN, and SSD evaluated with different backbones.
D. Improvements of Localization: Recent localization improvements fall into bounding-box refinement and new loss functions for more accurate localization.The paper identifies these as two groups of methods in recent detectors.

1) Bounding Box Refinement: … G. Weakly Supervised Object Detection

The reviewed methods improve object detection through bounding-box refinement, alternative localization losses, segmentation supervision, adversarial training, and weakly supervised learning. These approaches address localization accuracy, small or occluded objects, computational trade-offs, and reliance on bounding-box annotations.

1) Bounding Box Refinement:: Bounding-box refinement iteratively applies a BB regressor to detection results to improve predicted location and size, but may not monotonically improve or can degrade localization.The passage describes refinement as post-processing and notes convergence-oriented iteration alongside its potential failure modes.
2) New Loss Functions for Accurate Localization:: Traditional coordinate regression can mismatch final localization evaluation and lacks localization confidence, complicating non-maximum suppression for overlapping boxes.These drawbacks are especially evident for objects with very large aspect ratios.
E. Learning with Segmentation Loss: Segmentation-based learning improves detection by using semantic segmentation losses alongside detection.The passage presents object detection and semantic segmentation as fundamental computer-vision tasks whose losses can be combined for improvement.
E. Learning with Segmentation Loss: A segmentation network can serve as a fixed auxiliary feature extractor, providing an easy implementation at the cost of additional computation.The segmentation network is integrated into the detector as auxiliary features.
E. Learning with Segmentation Loss: An auxiliary segmentation branch can be trained with multi-task segmentation-and-detection losses, removed during inference, and therefore avoid affecting detection speed.This strategy requires pixel-level image annotations during training.
F. Adversarial Training: Adversarial training has been applied to improve detection of small and occluded objects by enhancing small-object features or generating occlusion masks.For small objects, GANs narrow representations between small and large objects; for occlusion, adversarial training generates masks.
G. Weakly Supervised Object Detection: Weakly supervised object detection reduces reliance on manually labeled bounding boxes by training detectors with image-level annotations.WSOD is motivated by the large amount of manually labeled data usually required for deep-learning-based detectors.
G. Weakly Supervised Object Detection: WSOD commonly formulates object candidates as a labeled bag in a multi-instance learning process, while other methods use class activation mapping, proposal ranking, masking, or generative adversarial training.Class activation mapping enables localization from image-level labels, and masking identifies likely objects when detection scores drop sharply.

H. Detection with Domain Adaptation · V. CONCLUSION AND FUTURE DIRECTIONS

The paper identifies domain adaptation as a way to narrow gaps caused by non-i.i.d. data and outlines future directions spanning efficient, end-to-end, small-object, 3D, video, and open-world detection.

H. Detection with Domain Adaptation: Domain adaptation offers a way to narrow the gap between domains when object detection encounters non-i.i.d. data.Most detectors are trained as likelihood-estimation systems assuming independent and identically distributed data.
V. CONCLUSION AND FUTURE DIRECTIONS: The survey reviews two decades of milestone detectors, key technologies, speed-up methods, datasets, and metrics while providing a roadmap toward future directions.Its stated aim is to help readers understand the fast-moving evolution of object detection.
V. CONCLUSION AND FUTURE DIRECTIONS: Lightweight detection targets faster inference on low-power edge devices, but speed gaps remain, especially for small objects and multi-source information.Applications include mobile augmented reality, autonomous driving, smart cities, smart cameras, and face verification.
V. CONCLUSION AND FUTURE DIRECTIONS: End-to-end detection seeks pipelines that combine one-to-one label assignment with high detection accuracy and efficiency.Most existing methods still use one-to-many assignment with separately designed non-maximum suppression.
V. CONCLUSION AND FUTURE DIRECTIONS: Small-object detection remains challenging in large scenes, motivating visual attention mechanisms and high-resolution lightweight networks.Applications include population or animal counting and military-target detection from satellite images.
V. CONCLUSION AND FUTURE DIRECTIONS: Video detection should exploit spatial and temporal correlations for real-time detection and tracking in HD videos under computational limits.Traditional detectors are generally designed for image-wise detection and ignore correlations between video frames.
V. CONCLUSION AND FUTURE DIRECTIONS: Open-world detection includes out-of-domain generalization, zero-shot detection, and incremental detection, with efforts addressing catastrophic forgetting and supplemental information.The passage contrasts these challenges with humans’ ability to discover unknown categories and learn new labeled knowledge while retaining existing patterns.