Glanz: aiMOOC über GPT aiMOOC Action erstellt

2026-09-01T05:08:20Z

aiMOOC über GPT aiMOOC Action erstellt

Neue Seite

{{T}}
[[Category:English]]
[[Category:Computer Vision]]

{{BR}}
= Introduction =

'''Computer vision''' is the study of methods that enable computers to extract useful information from images and video. It draws on [[English:Digital image processing|image processing]], [[English:Machine learning|machine learning]], [[English:Artificial intelligence|artificial intelligence]], geometry, optimization, statistics, and signal processing. A computer vision system may classify an entire image, locate objects, assign a class to every pixel, estimate depth or motion, reconstruct three-dimensional structure, read text, or connect visual observations with language.

At university level, computer vision is best understood as a chain of modeling decisions. You must decide how visual data are represented, which assumptions connect measurements to the physical world, what a model should predict, which loss function and evaluation metric are appropriate, and how the system will behave outside the training distribution. Modern systems often learn powerful representations from data, but classical geometry and image-processing ideas remain essential for calibration, feature matching, 3D reconstruction, robustness analysis, and debugging.

[[File:Convolutional Neural Network.png|500px|frameless|center]]

The diagram above introduces a common learning-based pipeline in which spatial feature extraction is followed by a prediction stage. Later sections compare this design with transformer-based approaches and geometric methods.

{{#ev:youtube|https://www.youtube.com/watch?v=2fq9wYslV0A|500|center}}

{{BR}}
== Learning Goals ==

After completing this aiMOOC, you should be able to explain how images are formed and represented, distinguish major computer vision tasks, reason about convolutional and transformer-based models, connect geometric vision with learned representations, select suitable metrics, diagnose dataset and distribution problems, and design a small end-to-end computer vision study with responsible evaluation.

{{BR}}
== Recommended Prerequisites ==

You will benefit from basic [[English:Linear algebra|linear algebra]], [[English:Calculus|calculus]], [[English:Probability|probability]], [[English:Python (programming language)|Python programming]], and introductory [[English:Machine learning|machine learning]]. Familiarity with vectors, matrices, derivatives, probability distributions, and gradient-based optimization will help you follow the mathematical ideas.

{{BR}}
= From Light to Digital Images =

A digital image is a sampled measurement of light. A grayscale image can be represented as a matrix, while a color image is often represented as a tensor with height, width, and color channels. Pixel values are measurements, not semantic facts: the same object can produce very different pixel values under changes in illumination, viewpoint, sensor response, focus, exposure, weather, or occlusion.

A useful geometric abstraction is the [[English:Pinhole camera model|pinhole camera model]]. In homogeneous coordinates, a 3D point can be projected to an image point by a camera matrix. A common form is <math>\mathbf{x} \sim \mathbf{K}[\mathbf{R}|\mathbf{t}]\mathbf{X}</math>, where <math>\mathbf{K}</math> contains intrinsic camera parameters, <math>\mathbf{R}</math> and <math>\mathbf{t}</math> describe camera pose, and the symbol <math>\sim</math> means equality up to scale. Real lenses add effects such as radial and tangential distortion, so calibration estimates parameters needed to relate image measurements to rays in 3D space.

[[File:Principles of imaging geometry of a camera.svg|500px|frameless|center]]

[[File:Multiple chessboard views.png|500px|frameless|center]]

Calibration patterns with known geometry make it possible to estimate camera parameters from multiple views. Calibration matters whenever image coordinates must be interpreted metrically, for example in [[English:Robotics|robotics]], [[English:Photogrammetry|photogrammetry]], autonomous systems, augmented reality, or 3D reconstruction.

{{BR}}
== Sampling, Color, and Preprocessing ==

Spatial resolution determines how densely a scene is sampled. Quantization determines how many intensity levels can be represented. Aliasing appears when high-frequency scene content is sampled too coarsely. Color can be represented in spaces such as RGB, HSV, Lab, or YCbCr, each of which can be useful for different processing goals.

Preprocessing may include resizing, normalization, denoising, contrast adjustment, geometric rectification, or augmentation. However, every transformation changes the data distribution. You should treat preprocessing as part of the model pipeline and document it so that training, validation, testing, and deployment use compatible assumptions.

{{BR}}
= Classical Visual Features and Geometry =

Before deep learning became dominant, many computer vision pipelines used hand-designed local features. These methods remain important because they expose the structure of visual reasoning and still perform well in settings where geometry, efficiency, or limited data matter.

Edges are locations of strong intensity change. Corners are image points with distinctive local variation in multiple directions. The Harris corner detector, for example, uses local image gradients to find points whose neighborhoods change strongly under small shifts. Such points can serve as repeatable landmarks across views.

[[File:Harris corners detected on chessboard.png|500px|frameless|center]]

Descriptors such as [[English:Scale-invariant feature transform|SIFT]] summarize local appearance in a way designed to be robust to moderate changes in scale, rotation, and illumination. Feature matching then proposes correspondences between images. Robust estimators such as [[English:Random sample consensus|RANSAC]] can separate geometrically consistent matches from outliers.

[[File:SIFT gradient magnitude and orientation computation.svg|500px|frameless|center]]

The [[English:Hough transform|Hough transform]] uses a voting process in a parameter space to detect structures such as lines or circles. It demonstrates a broader idea in vision: a difficult grouping problem in image space can become easier when transformed into a representation where evidence accumulates for shared parameters.

[[File:Hough transform diagram.svg|500px|frameless|center]]

{{BR}}
== Multiple-View Geometry ==

When the same scene is observed from different viewpoints, correspondences constrain camera motion and scene structure. The [[English:Epipolar geometry|epipolar constraint]] relates matching points in two views through the fundamental matrix. If camera intrinsics are known, the essential matrix represents the corresponding calibrated relation. Triangulation estimates 3D point positions from rays in multiple views.

A [[English:Homography|homography]] is a projective transformation between two images of a planar scene, or between views related by pure camera rotation. It is useful for panorama stitching, planar tracking, document rectification, and augmented-reality overlays. You should not use a single homography to model a general non-planar scene with substantial translation because parallax violates that assumption.

{{BR}}
= Learning from Visual Data =

In supervised learning, a model is trained from input-output pairs such as images and class labels, bounding boxes, segmentation masks, depth maps, or keypoints. The training objective usually minimizes an empirical loss plus regularization. A learned model is only as meaningful as its data, labels, objective, and evaluation protocol.

For image classification, a model maps an image to one or more category predictions. A linear classifier can already illustrate core ideas such as score functions, decision boundaries, softmax probabilities, loss functions, and regularization.

{{#ev:youtube|https://www.youtube.com/watch?v=pdqofxJeBN8|500|center}}

{{BR}}
== Convolutional Neural Networks ==

A [[English:Convolutional neural network|convolutional neural network]] applies learned filters across spatial locations. Local connectivity lets a filter respond to patterns in a neighborhood, while weight sharing lets the same filter detect a pattern at many positions. Stacking convolution, nonlinear activation, normalization, downsampling, and residual connections allows a network to learn increasingly abstract features.

For a two-dimensional discrete feature map <math>x</math> and kernel <math>w</math>, a cross-correlation operation commonly used in deep-learning libraries can be written as <math>y(i,j)=\sum_{u,v} w(u,v)x(i+u,j+v)</math>. Padding controls border handling, stride controls the movement of the kernel, and dilation changes the spacing between sampled input positions. The receptive field of deeper units grows as layers are composed.

[[File:Convolutional Network (vector).svg|500px|frameless|center]]

[[File:CNN Convolutional Layers.svg|500px|frameless|center]]

Training typically uses mini-batch stochastic optimization and backpropagation. Data augmentation, normalization, weight decay, learning-rate schedules, transfer learning, and careful validation can strongly affect performance. A high-capacity architecture cannot compensate for mislabeled data, data leakage, or an evaluation set that does not represent the intended deployment conditions.

{{#ev:youtube|https://www.youtube.com/watch?v=oGpzWAlP5p0|500|center}}

{{BR}}
== Attention and Vision Transformers ==

[[English:Transformer (deep learning architecture)|Transformers]] replace or complement convolution with attention mechanisms. In a vision transformer, an image is commonly divided into patches, each patch is mapped to an embedding, positional information is added, and self-attention mixes information across tokens. Attention can model long-range interactions directly, while convolution builds in strong local spatial biases.

Neither architecture is universally superior. Performance depends on scale, data, pretraining, computation, latency, memory, and the task. Hybrid systems combine convolution, attention, multiscale feature hierarchies, or pretrained visual encoders. Modern vision systems are often selected by benchmarking under the constraints of the actual application rather than by architecture name alone.

{{#ev:youtube|https://www.youtube.com/watch?v=RQowiOF_FvQ|500|center}}

{{BR}}
= Major Computer Vision Tasks =

{{BR}}
== Image Classification ==

Image classification predicts a label for an entire image. In single-label classification, top-1 accuracy is common, but it can hide class imbalance. Macro-averaged precision, recall, or F1 can be more informative when rare classes matter. Calibration of predicted probabilities is a separate property from classification accuracy.

{{BR}}
== Object Detection ==

Object detection predicts object categories and spatial regions, often bounding boxes. Detectors may use one-stage or two-stage designs. Training must solve both classification and localization. Evaluation commonly uses [[English:Intersection over union|intersection over union]] to compare predicted and reference boxes and average precision to summarize precision-recall behavior across confidence thresholds.

[[File:ObjectDetectionExample.png|500px|frameless|center]]

{{BR}}
== Semantic, Instance, and Panoptic Segmentation ==

Semantic segmentation assigns a class to each pixel. Instance segmentation additionally separates distinct object instances. Panoptic segmentation combines semantic labeling of background regions with instance-level labeling of countable objects.

[[File:Image-segmentation-example.jpg|500px|frameless|center]]
[[File:Image-segmentation-example-segmented.png|500px|frameless|center]]
[[File:Mask R-CNN.webp|500px|frameless|center]]

For segmentation, pixel accuracy can be misleading when one class dominates. Intersection over union, mean IoU across classes, Dice score, and class-specific metrics are often more informative. Evaluation should report which classes are difficult and whether errors occur at object boundaries, in small objects, or under occlusion.

{{#ev:youtube|https://www.youtube.com/watch?v=PTypu6GqEd4|500|center}}

{{BR}}
== Motion and Optical Flow ==

[[English:Optical flow|Optical flow]] estimates apparent motion between image frames. Classical formulations often rely on brightness constancy and local smoothness assumptions, while modern learned approaches can estimate flow from data. Occlusion, motion blur, textureless regions, large displacements, and illumination changes make the problem difficult.

[[File:Optical flow example v2.png|500px|frameless|center]]

Motion estimates support tracking, video stabilization, action analysis, navigation, video compression, and scene understanding. When motion is interpreted geometrically, you must distinguish image-plane motion from physical 3D motion.

{{BR}}
== Depth, Pose, and 3D Reconstruction ==

Depth can be estimated from stereo correspondence, multiple views, active sensors, monocular learned priors, or combinations of these sources. Camera pose estimation and structure from motion recover camera trajectories and scene geometry from image correspondences. In 3D computer vision, uncertainty matters: repeated textures, reflective surfaces, low texture, moving objects, and calibration errors can produce ambiguous or unstable reconstructions.

{{BR}}
== Vision and Language ==

Vision-language systems connect visual representations with text for tasks such as image captioning, visual question answering, retrieval, grounding, and multimodal assistance. These systems can reuse large pretrained encoders and learn joint representations across modalities.

[[File:Image-captioning.png|500px|frameless|center]]

Because natural-language outputs can sound confident even when visual evidence is weak, evaluation should distinguish grounded visual recognition from plausible language generation. In safety-relevant settings, the system should expose uncertainty and support human verification.

{{BR}}
= Datasets, Training, and Generalization =

A dataset defines what a model can learn and what an experiment can claim. You should separate training, validation, and test data so that model selection does not use information from the final test set. Near-duplicate images, frames from the same video, or samples from the same person can cause leakage across splits and inflate apparent performance.

Class balance, long-tail distributions, annotation quality, background correlations, demographic representation, camera type, geography, and capture conditions can all affect outcomes. A model may learn shortcuts, such as recognizing a background rather than the intended object. Data augmentation can improve invariance, but unrealistic augmentation can also create artifacts or hide deployment-specific failure modes.

Transfer learning starts from a pretrained model and adapts it to a new task. Self-supervised learning learns representations from unlabeled or weakly labeled data. Few-shot and zero-shot methods attempt to generalize with limited task-specific labels. These strategies reduce annotation requirements, but they do not remove the need to test domain shift and failure cases.

{{BR}}
== Experimental Design ==

A strong experiment begins with a falsifiable question. Define the dataset split before tuning, select baselines that isolate the value of the proposed method, report uncertainty when feasible, and avoid choosing metrics only because they make a result look favorable. Ablation studies test which components actually contribute. Error analysis groups failures into meaningful categories and often reveals more than a single aggregate score.

When comparing models, keep training budget, input resolution, augmentation, pretraining, and evaluation protocol as controlled as possible. If resources differ, report them explicitly. Efficiency metrics such as latency, throughput, memory use, energy consumption, and model size can be as important as predictive accuracy.

{{BR}}
= Evaluation Metrics and Failure Analysis =

For a binary classifier, [[English:Precision and recall|precision]] measures the fraction of predicted positives that are correct, while recall measures the fraction of actual positives that are found. The F1 score is their harmonic mean. For detection and segmentation, overlap-based metrics such as IoU are common. For depth, error can be reported in absolute, relative, or scale-aware forms. For tracking, metrics may combine identity consistency and localization quality.

No metric is neutral. A medical triage system may value recall differently from an automated quality-control system. A robot may need low latency and calibrated uncertainty. A fairness-sensitive application may require subgroup analysis. You should therefore connect every metric to the costs of errors in the intended use.

A confusion matrix can reveal class-specific mistakes. Confidence distributions can reveal overconfidence. Slice-based evaluation tests performance under conditions such as low light, motion blur, rain, unusual viewpoints, rare classes, or different devices. Stress testing should include both expected use and plausible misuse.

{{BR}}
= Robustness, Ethics, and Responsible Deployment =

Computer vision systems can affect privacy, safety, accessibility, employment, policing, medicine, transport, and public space. Responsible design therefore requires more than achieving high benchmark scores.

You should ask who is represented in the data, who may be harmed by errors, whether people gave meaningful consent for data collection, whether the system enables intrusive surveillance, and whether performance varies across relevant groups or environments. Face analysis, biometric identification, and behavioral inference deserve particularly careful legal and ethical scrutiny because they can affect people who did not choose to participate.

Robustness includes sensitivity to corruptions, adversarial perturbations, camera changes, domain shift, and previously unseen scenarios. A deployment plan should define monitoring, fallback behavior, human oversight, data retention, versioning, and procedures for handling incidents. A model that is accurate on average can still be unsafe if its errors cluster in high-consequence situations.

[[File:Night run reconstruction.png|500px|frameless|center]]

The event-camera reconstruction above is a useful reminder that sensing hardware and algorithms interact. Better visual inference can come from changing the sensor, the representation, the model, or all three.

{{BR}}
= Practical Workflow for a Vision Project =

A reproducible university project can follow a compact cycle: formulate a task and success criteria, inspect data and licensing, build a simple baseline, choose metrics, train under a fixed protocol, perform error analysis, test robustness, document limitations, and package code and model metadata so another person can reproduce the result.

Keep a clear distinction between development performance and final test performance. Save random seeds and environment information where possible. Record preprocessing and augmentation. Use version control for code and configuration. For models used outside a classroom, include documentation describing intended use, excluded uses, important failure modes, evaluation coverage, and maintenance responsibilities.

{{BR}}
= Interactive Tasks =

{{BR}}
== Quiz: Test Your Knowledge ==

{{MC}}

'''What does camera calibration primarily estimate?'''
(Parameters that relate image measurements to camera geometry)
(!The semantic class of every pixel)
(!The learning rate of a neural network)
(!The number of objects in every image)

{{E}}
<br>

{{MC}}

'''Which property of convolution enables one learned filter to be applied across many image locations?'''
(Weight sharing)
(!Random cropping)
(!Label smoothing)
(!Early stopping)

{{E}}
<br>

{{MC}}

'''Which task assigns a semantic class to each pixel?'''
(Semantic segmentation)
(!Image classification)
(!Image retrieval)
(!Camera calibration)

{{E}}
<br>

{{MC}}

'''What does intersection over union measure?'''
(Overlap between two spatial regions)
(!Similarity between two learning rates)
(!Distance between two color channels)
(!Number of layers in a model)

{{E}}
<br>

{{MC}}

'''Which method is commonly used to fit a model despite many outlier correspondences?'''
(RANSAC)
(!Softmax)
(!Dropout)
(!Pooling)

{{E}}
<br>

{{MC}}

'''What is a key limitation of a single homography for a general three dimensional scene with camera translation?'''
(Parallax violates the planar mapping assumption)
(!It cannot represent any rotation)
(!It requires grayscale images)
(!It prevents feature extraction)

{{E}}
<br>

{{MC}}

'''What is the central operation used by a vision transformer to mix information across image tokens?'''
(Self attention)
(!Histogram equalization)
(!Nearest neighbor interpolation)
(!Background subtraction)

{{E}}
<br>

{{MC}}

'''Why can test leakage make a computer vision model appear better than it is?'''
(Information from evaluation data influences model development)
(!The camera uses too many pixels)
(!The images contain color channels)
(!The optimizer uses gradients)

{{E}}
<br>

{{MC}}

'''Which statement best describes optical flow?'''
(Apparent image motion between frames)
(!A list of object class names)
(!A camera lens distortion parameter)
(!A probability calibration curve)

{{E}}
<br>

{{MC}}

'''Why should computer vision evaluation include subgroup and condition specific analysis?'''
(Aggregate metrics can hide concentrated failure modes)
(!All images have identical difficulty)
(!Every model has the same confidence)
(!Training data always match deployment data)

{{E}}
<br>

{{BR}}
== Memory Game ==

<div class="memo-quiz">
{|
|-
| Pixel || Sampled image measurement at a spatial location
|-
| Convolution || Local filter operation reused across an image
|-
| Homography || Projective mapping for planar scenes or pure camera rotation
|-
| Triangulation || Estimation of a 3D point from multiple viewing rays
|-
| IoU || Ratio of intersection area to union area for two regions
|-
| Occlusion || Situation in which one scene element blocks another from view
|}
{{E}}
<br>

{{BR}}
== Drag and Drop ==

<div class="lueckentext-quiz">

{| class="wikitable"
! Match the correct terms.
! Topic
|-
| '''Image classification'''
| Assign one category to an entire image
|-
| '''Object detection'''
| Predict categories together with spatial boxes
|-
| '''Semantic segmentation'''
| Assign a semantic category to each pixel
|-
| '''Optical flow'''
| Estimate apparent motion between video frames
|-
| '''Camera calibration'''
| Estimate parameters that connect image measurements with camera geometry
|}
{{E}}

<br>
...
<br />

{{BR}}
== Crossword Puzzle ==

<div class="kreuzwort-quiz">
{|
|-
| Calibration || What process estimates camera parameters from known geometric observations?
|-
| Segmentation || What task divides an image into meaningful labeled regions?
|-
| Convolution || What local filtering operation is central to many CNNs?
|-
| Homography || What projective mapping relates views of a plane?
|-
| Transformer || What architecture uses self attention over visual tokens?
|-
| Occlusion || What occurs when one object blocks another from the camera?
|}
{{E}}
<br>

{{BR}}
== LearningApps ==

<iframe> https://learningapps.org/index.php?s=Computer+Vision </iframe>

{{BR}}
== Cloze Text ==

<quiz display=simple>
{'''Complete the text.'''<br>
|type="{}"}
A digital image stores sampled visual measurements in an array of { pixels }. Camera calibration estimates parameters that connect image coordinates with { geometry }. A convolutional layer learns spatial filters through shared { weights }. Object detection combines category prediction with spatial { localization }. Semantic segmentation predicts a label for each { pixel }. Optical flow estimates apparent motion between successive { frames }. A vision transformer mixes patch representations using self { attention }. Reliable evaluation requires a held out { test } set that is not used for tuning. Responsible deployment requires analysis of uncertainty, privacy, bias, and possible { harms }.
</quiz>

<br>

{{BR}}
= Open-Ended Tasks =

{{BR}}
=== Easy ===
# [[English:Visual dataset audit|Visual dataset audit]]: Choose a small public image dataset, inspect at least fifty samples, and create a one-page illustrated report describing labels, image quality, class balance, and possible sources of bias.
# [[English:Edge detection experiment|Edge detection experiment]]: Apply two edge-detection settings to the same set of images, save representative outputs, and explain how threshold choices change missed and false edges.
# [[English:Computer vision interview|Computer vision interview]]: Interview a student, researcher, technician, or developer who uses images or video in their work and write a short summary of the problem, data, evaluation criteria, and failure cases they consider important.
# [[English:Vision concept video|Vision concept video]]: Produce a two-minute explainer video that teaches one concept such as convolution, IoU, calibration, or segmentation using your own diagrams and one concrete example.

{{BR}}
=== Standard ===
# [[English:Classifier baseline study|Classifier baseline study]]: Train a simple image classifier on a small dataset, compare it with a transfer-learning baseline, and report accuracy, a confusion matrix, training time, and at least three error categories.
# [[English:Feature matching investigation|Feature matching investigation]]: Capture overlapping photographs of a planar scene, extract and match local features, estimate a homography with robust fitting, visualize inliers and outliers, and explain when the mapping fails.
# [[English:Segmentation annotation project|Segmentation annotation project]]: Design a labeling guide for a three-class segmentation problem, annotate a small image set with a partner, measure disagreement, revise the guide, and discuss annotation ambiguity.
# [[English:Computer vision field visit|Computer vision field visit]]: Visit a laboratory, makerspace, robotics facility, medical-imaging unit, production site, museum technology exhibit, or other relevant location and document how sensing, lighting, calibration, and human oversight influence a real visual system.

{{BR}}
=== Advanced ===
# [[English:Robustness benchmark|Robustness benchmark]]: Build a controlled evaluation set with variations such as blur, low light, compression, viewpoint change, or occlusion, test a pretrained vision model, and analyze how performance changes across conditions.
# [[English:Detection error taxonomy|Detection error taxonomy]]: Evaluate an object detector on a public dataset, categorize false positives and false negatives by cause, relate the categories to IoU and confidence thresholds, and propose targeted improvements.
# [[English:Vision architecture research poster|Vision architecture research poster]]: Compare a convolutional model and a transformer-based model under a matched experimental budget, then create a research poster covering method, compute, metrics, uncertainty, and limitations.
# [[English:Responsible vision deployment plan|Responsible vision deployment plan]]: Design a deployment proposal for a high-impact computer vision application that includes stakeholders, data governance, subgroup evaluation, privacy safeguards, uncertainty handling, human escalation, monitoring, and a stop-use criterion.

{{:Open Task - Create a MOOC}}

{{BR}}
= Learning Assessment =

# [[English:Metric selection case study|Metric selection case study]]: Given a vision application with asymmetric error costs, justify a set of primary and secondary metrics and explain why accuracy alone would be insufficient.
# [[English:Geometry transfer problem|Geometry transfer problem]]: Analyze whether a homography, epipolar model, or direct learned predictor is appropriate for three different camera and scene configurations, and defend each choice from the stated assumptions.
# [[English:Model comparison report|Model comparison report]]: Compare two models using predictive quality, latency, memory, training data, and robustness results, then recommend one for a constrained deployment and state what evidence is still missing.
# [[English:Dataset shift analysis|Dataset shift analysis]]: Examine a scenario in which training images and deployment images differ, identify likely shortcut features and failure modes, and design a test set that could reveal them.
# [[English:Error analysis defense|Error analysis defense]]: Present five representative model failures, group them into a defensible taxonomy, connect each group to a possible cause, and propose one experiment that could distinguish competing explanations.
# [[English:Responsible design review|Responsible design review]]: Evaluate a proposed computer vision system for privacy, bias, misuse, uncertainty, and human oversight, then decide whether deployment should proceed, be modified, or be rejected.

<br>
<br>

{{BR}}
= Evidence of Learning =

'''Knowledge:''' You can explain image formation, sampling, calibration, local features, convolution, attention, detection, segmentation, optical flow, 3D vision, and the logic behind commonly used evaluation metrics.

'''Skills:''' You can prepare visual data, implement or adapt a baseline, select an evaluation protocol, visualize predictions, perform quantitative and qualitative error analysis, compare alternatives fairly, and communicate uncertainty.

'''Products:''' Strong evidence includes reproducible notebooks or code, dataset documentation, experiment logs, figures, model cards or equivalent documentation, research posters, short videos, and written analyses that connect results to assumptions.

'''Transfer:''' You can recognize when a technique is inappropriate, adapt a method to a new domain, detect likely dataset leakage or shortcut learning, select metrics based on real costs, and propose safeguards for responsible deployment.

<br>
<br>

{{BR}}
= OERs on the Topic =

The English Wikipedia article below provides a broad open reference for terminology, history, applications, and related methods.

<iframe> https://en.m.wikipedia.org/wiki/Computer_vision </iframe>

<br>

{{BR}}
= Linked Learning Areas =

{| align=center
{{:D-Tab}}
'''[[English:Computer vision|Computer vision]]'''
# [[English:Digital image processing|Digital image processing]]
# [[English:Machine learning|Machine learning]]
# [[English:Artificial intelligence|Artificial intelligence]]
# [[English:Convolutional neural network|Convolutional neural network]]
# [[English:Transformer (deep learning architecture)|Transformer (deep learning architecture)]]
# [[English:Object detection|Object detection]]
# [[English:Image segmentation|Image segmentation]]
# [[English:Optical flow|Optical flow]]
# [[English:Multiple view geometry|Multiple view geometry]]
# [[English:Robotics|Robotics]]
# [[English:Photogrammetry|Photogrammetry]]
# [[English:Responsible artificial intelligence|Responsible artificial intelligence]]
|}

{{BR}}
= aiMOOC Projects =
[[Category:English]]
[[Category:Computer Vision]]
[[Category:Higher Education]]
[[Category:Computer Science]]
[[Category:Artificial Intelligence]]
[[Category:Machine Learning]]
[[Category:Image Processing]]
[[Category:Robotics]]
[[Category:AI_MOOC]]
[[Category:GPT aiMOOC]]
{{MT}}

English:Computer Vision - Versionsgeschichte

Glanz: aiMOOC über GPT aiMOOC Action erstellt