COMPUTER VISION + YOLO · 3 LEVELS

Teach AIto detect objects

Collect objects, draw boxes, train YOLO, and test it live with a webcam.

CIT independently designed this course from the public Ultralytics repository and official documentation. It is not an official, certified, or jointly developed Ultralytics course.

Students across three age groups experimenting with object detection — OBJECT-ONLY CAMERA LAB

3grade-band courses

8 · 12 · 16sessions by level

1:1 · SMALL GROUPclass format

ONLINE · IN PERSONhow we run it

WEBCAMhands-on now

TELLO · ROBOMASTERfuture extension

SEE THE SYSTEM

Five roles working as one AI system

Every lesson connects Input, Memory, Process, Output, and Control as one unit. Control checks the result, changes one input, stored rule, or process step, and runs the system again.

Input

An object photo or webcam frame enters the system.

Explore the five roles

A student testing webcam object detection — An instructional stage visualization, not a real model score.

EXPLORE THE PIPELINE

From one photo to a detection

This explorable diagram connects collection, labels, splits, training, validation, webcam use, and the safety boundary for a future robot.

Open full screen ↗

CHOOSE YOUR DEPTH

One YOLO workflow, three levels of depth

Object Detectives: YOLO Computer Vision

Use an object detector, label two or three classroom objects, and train a small YOLO model with tutor support.

Format: 8 × 90 min
Prerequisites: Students can begin if they can open and save a file and compare two rows in a table. No Python experience is required. The tutor runs training commands while students explain data, predictions, and errors. The webcam points only at toys and classroom objects, never faces or private spaces.
Milestone: Object data card, before-and-after comparison, webcam Object Detective demo, and a child-friendly model card

01A camera turns the world into numbers

What is different when a person sees a toy car and when a computer receives a photo?

Evidence

A four-part card separating human descriptions from pixel input

InputData in
The lesson question and a fixed non-personal example: What is different when a person sees a toy car and when a computer receives a photo?
MemoryStored state
The rules and intermediate record kept for the next step: An observation board showing one object under different light, distance, and direction
ProcessTransform
Compare, build, and test with these ideas: pixel, image, input, computer vision.
OutputResult
A four-part card separating human descriptions from pixel input
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

02Naming an object versus locating it

How is saying that a car is present different from locating it in the image?

Evidence

A visual explanation comparing classification and detection

InputData in
The lesson question and a fixed non-personal example: How is saying that a car is present different from locating it in the image?
MemoryStored state
The rules and intermediate record kept for the next step: A detection sheet for reading boxes, labels, and confidence from a pretrained model
ProcessTransform
Compare, build, and test with these ideas: classification, object detection, bounding box, confidence.
OutputResult
A visual explanation comparing classification and detection
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

03A useful photo set has variety

Why are many front photos of the same blue car not enough?

Evidence

An object data card stating what will and will not be photographed

InputData in
The lesson question and a fixed non-personal example: Why are many front photos of the same blue car not enough?
MemoryStored state
The rules and intermediate record kept for the next step: A safe photo plan covering angle, size, background, and occlusion for two or three objects
ProcessTransform
Compare, build, and test with these ideas: dataset, variation, background, data card.
OutputResult
An object data card stating what will and will not be photographed
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

04Draw answer boxes with one consistent rule

What answer does the model learn if only half of an object is inside the box?

Evidence

A small annotation set checked by a partner against three labeling rules

InputData in
The lesson question and a fixed non-personal example: What answer does the model learn if only half of an object is inside the box?
MemoryStored state
The rules and intermediate record kept for the next step: A set of consistently drawn answer boxes in the provided labeling tool
ProcessTransform
Compare, build, and test with these ideas: annotation, class label, box consistency, quality check.
OutputResult
A small annotation set checked by a partner against three labeling rules
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

05Separate practice photos from test photos

Why is it not a fair test if the model has already seen the test photo?

Evidence

A split map explaining the role of train, validation, and test sets

InputData in
The lesson question and a fixed non-personal example: Why is it not a fair test if the model has already seen the test photo?
MemoryStored state
The rules and intermediate record kept for the next step: A color-card split that keeps one photo burst from leaking across sets
ProcessTransform
Compare, build, and test with these ideas: train, validation, test, data leakage.
OutputResult
A split map explaining the role of train, validation, and test sets
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

06YOLO improves predictions through repetition

What changes after YOLO compares a wrong prediction with the answer?

Evidence

A learning-loop diagram connecting prediction, answer, difference, and weight update

InputData in
The lesson question and a fixed non-personal example: What changes after YOLO compares a wrong prediction with the answer?
MemoryStored state
The rules and intermediate record kept for the next step: A comparison board saving early and later predictions from a tutor-run nano model training
ProcessTransform
Compare, build, and test with these ideas: pretrained model, epoch, loss, weight.
OutputResult
A learning-loop diagram connecting prediction, answer, difference, and weight update
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

07Test new objects with a webcam

What should we test next if an object is found on a bright desk but missed on a dark background?

Evidence

A webcam test sheet containing both successful and failed cases

InputData in
The lesson question and a fixed non-personal example: What should we test next if an object is found on a bright desk but missed on a dark background?
MemoryStored state
The rules and intermediate record kept for the next step: A webcam object test changing light, distance, and occlusion one at a time
ProcessTransform
Compare, build, and test with these ideas: inference, webcam frame, false positive, miss.
OutputResult
A webcam test sheet containing both successful and failed cases
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

08Object Detectives explain model limits

How can we show what our model does well and what should not yet be trusted?

Evidence

A child-friendly model card and a three-minute live demonstration

InputData in
The lesson question and a fixed non-personal example: How can we show what our model does well and what should not yet be trusted?
MemoryStored state
The rules and intermediate record kept for the next step: An Object Detective exhibit connecting photos, label rules, the training loop, and webcam tests
ProcessTransform
Compare, build, and test with these ideas: model card, demo, evidence, limitation.
OutputResult
A child-friendly model card and a three-minute live demonstration
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

YOLO Vision Builders

Read the official GitHub repository, then run, train, and evaluate a custom three-class detector.

Format: 12 × 105 min
Prerequisites: Students should be able to read Python variables, conditionals, loops, and run one terminal command. Students with less coding experience begin from a provided notebook and fixed dataset. Datasets exclude faces, name tags, and private spaces, and webcam work uses an object-only zone.
Milestone: A reproducible GitHub repository, custom YOLO weights, error analysis, webcam demo, and model card

01Narrow the computer vision problem

Why does object detection fit our mission better than classification or segmentation?

Evidence

A three-class detection problem statement with success and failure examples

InputData in
The lesson question and a fixed non-personal example: Why does object detection fit our mission better than classification or segmentation?
MemoryStored state
The rules and intermediate record kept for the next step: A problem canvas comparing the inputs and outputs of three vision tasks
ProcessTransform
Compare, build, and test with these ideas: task, classification, detection, segmentation.
OutputResult
A three-class detection problem statement with success and failure examples
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

02Reproduce pretrained YOLO first

What problems can an official example expose before we train anything?

Evidence

A reproduction note recording command, version, input, and output path

InputData in
The lesson question and a fixed non-personal example: What problems can an official example expose before we train anything?
MemoryStored state
The rules and intermediate record kept for the next step: A baseline run of one model on a fixed image and short video
ProcessTransform
Compare, build, and test with these ideas: installation, pretrained weights, predict, Results.
OutputResult
A reproduction note recording command, version, input, and output path
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

03Read the GitHub repository as a tool map

Which questions are answered by the README, docs, source, issues, and license?

Evidence

A pinned commit link and current-documentation checklist

InputData in
The lesson question and a fixed non-personal example: Which questions are answered by the README, docs, source, issues, and license?
MemoryStored state
The rules and intermediate record kept for the next step: A repository map linking five sections to the questions they answer
ProcessTransform
Compare, build, and test with these ideas: repository, commit, README, issue, license.
OutputResult
A pinned commit link and current-documentation checklist
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

04Define classes and edge cases first

What rule is needed when two similar objects belong to different classes?

Evidence

A dataset plan with class names, definitions, and boundary cases

InputData in
The lesson question and a fixed non-personal example: What rule is needed when two similar objects belong to different classes?
MemoryStored state
The rules and intermediate record kept for the next step: A class rule sheet listing included, excluded, and ambiguous examples
ProcessTransform
Compare, build, and test with these ideas: class ontology, edge case, in scope, out of scope.
OutputResult
A dataset plan with class names, definitions, and boundary cases
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

05Collect across backgrounds and occlusion

How can we test why condition diversity matters more than raw photo count?

Evidence

A data card with condition coverage and exclusion reasons

InputData in
The lesson question and a fixed non-personal example: How can we test why condition diversity matters more than raw photo count?
MemoryStored state
The rules and intermediate record kept for the next step: An object-only image set covering light, background, distance, direction, and occlusion
ProcessTransform
Compare, build, and test with these ideas: coverage, lighting, scale, occlusion, privacy.
OutputResult
A data card with condition coverage and exclusion reasons
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

06Audit YOLO labels and box quality

Why are box coordinates stored as numbers between zero and one?

Evidence

A QA log checking box tightness, class IDs, misses, and extra labels

InputData in
The lesson question and a fixed non-personal example: Why are box coordinates stored as numbers between zero and one?
MemoryStored state
The rules and intermediate record kept for the next step: An annotation audit that reconstructs one label row as a box
ProcessTransform
Compare, build, and test with these ideas: normalized coordinate, class id, annotation QA, empty image.
OutputResult
A QA log checking box tightness, class IDs, misses, and extra labels
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

07Create splits without leakage

Why can near-duplicate frames across train and test inflate the score?

Evidence

A dataset YAML with split rules and duplicate-check results

InputData in
The lesson question and a fixed non-personal example: Why can near-duplicate frames across train and test inflate the score?
MemoryStored state
The rules and intermediate record kept for the next step: Train, validation, and test directories grouped by capture session
ProcessTransform
Compare, build, and test with these ideas: group split, near duplicate, leakage, held-out test.
OutputResult
A dataset YAML with split rules and duplicate-check results
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

08Train a custom nano model

Why begin from pretrained weights instead of training from scratch?

Evidence

best.pt, results plot, configuration sheet, and rerun command

InputData in
The lesson question and a fixed non-personal example: Why begin from pretrained weights instead of training from scratch?
MemoryStored state
The rules and intermediate record kept for the next step: A first custom training run recording version, model, data YAML, epochs, and image size
ProcessTransform
Compare, build, and test with these ideas: transfer learning, epoch, batch, checkpoint.
OutputResult
best.pt, results plot, configuration sheet, and rerun command
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

09Connect precision, recall, and mAP to visible errors

Which metrics distinguish a model with many false alarms from one that misses objects?

Evidence

An evaluation card with per-class metrics and representative error images

InputData in
The lesson question and a fixed non-personal example: Which metrics distinguish a model with many false alarms from one that misses objects?
MemoryStored state
The rules and intermediate record kept for the next step: An evaluation board linking test metrics to visible false positives and false negatives
ProcessTransform
Compare, build, and test with these ideas: precision, recall, IoU, mAP, confusion matrix.
OutputResult
An evaluation card with per-class metrics and representative error images
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

10Change one variable to reduce errors

Can we know what helped if data, threshold, and epochs all change together?

Evidence

An experiment record with hypothesis, one variable, fixed conditions, result, and counterexample

InputData in
The lesson question and a fixed non-personal example: Can we know what helped if data, threshold, and epochs all change together?
MemoryStored state
The rules and intermediate record kept for the next step: A before-and-after experiment on one slice such as occlusion or dark backgrounds
ProcessTransform
Compare, build, and test with these ideas: error slice, one-variable experiment, baseline, comparison.
OutputResult
An experiment record with hypothesis, one variable, fixed conditions, result, and counterexample
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

11Balance webcam confidence and speed

How do false alarms and misses change when the confidence threshold rises?

Evidence

A demo sheet with threshold rationale, latency observations, and failure cases

InputData in
The lesson question and a fixed non-personal example: How do false alarms and misses change when the confidence threshold rises?
MemoryStored state
The rules and intermediate record kept for the next step: A live webcam demo comparing two thresholds in an object-only zone
ProcessTransform
Compare, build, and test with these ideas: webcam source, confidence threshold, latency, frame.
OutputResult
A demo sheet with threshold rationale, latency observations, and failure cases
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

12Package the project so another person can reproduce it

Can someone reproduce the result from code and best.pt alone?

Evidence

A reproducible repository, webcam demo, and five-minute technical explanation

InputData in
The lesson question and a fixed non-personal example: Can someone reproduce the result from code and best.pt alone?
MemoryStored state
The rules and intermediate record kept for the next step: A GitHub project documenting setup, data boundaries, execution, evaluation, failures, and license review
ProcessTransform
Compare, build, and test with these ideas: README, environment, data card, model card, license.
OutputResult
A reproducible repository, webcam demo, and five-minute technical explanation
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

YOLO Computer Vision Research Studio

Reproduce a custom detector from pinned sources and defend dataset leakage, mAP, error slices, ablations, latency, and licensing choices.

Format: 16 × 120 min
Prerequisites: Experience with Python functions and classes, virtual environments, Git commits, and table-based analysis is recommended. Students should be comfortable with ratios, coordinates, averages, and confusion matrices. Projects use object-only datasets and exclude identity recognition, biometric inference, and unconsented recording.
Milestone: A version-pinned repository, custom detector, experiment table, robustness report, webcam application, model card, and technical defense

01Pin the environment and source

What must be recorded to reproduce today's result on another computer a month later?

Evidence

A clean-environment installation verification and first commit

InputData in
The lesson question and a fixed non-personal example: What must be recorded to reproduce today's result on another computer a month later?
MemoryStored state
The rules and intermediate record kept for the next step: An environment manifest recording Python, Ultralytics, model, device, and source commit
ProcessTransform
Compare, build, and test with these ideas: virtual environment, package pin, commit SHA, reproducibility.
OutputResult
A clean-environment installation verification and first commit
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

02Separate public APIs from internal implementation

How should we separate stable CLI and Python APIs from internal code that may change?

Evidence

A change-risk table for CLI, Python API, configuration, and internal source

InputData in
The lesson question and a fixed non-personal example: How should we separate stable CLI and Python APIs from internal code that may change?
MemoryStored state
The rules and intermediate record kept for the next step: A repository map linking public entry points in the README and docs to source folders
ProcessTransform
Compare, build, and test with these ideas: public API, CLI, Python API, configuration, internal module.
OutputResult
A change-risk table for CLI, Python API, configuration, and internal source
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

03Turn pretrained inference into a measurable baseline

What should be logged to compare predictions and latency across inputs instead of showing one polished demo?

Evidence

A baseline table covering input, device, image size, confidence, warm-up, and results

InputData in
The lesson question and a fixed non-personal example: What should be logged to compare predictions and latency across inputs instead of showing one polished demo?
MemoryStored state
The rules and intermediate record kept for the next step: A baseline script saving prediction output and timing on a fixed image set
ProcessTransform
Compare, build, and test with these ideas: inference, confidence, latency, warm-up, batch.
OutputResult
A baseline table covering input, device, image size, confidence, warm-up, and results
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

04Approve the problem, baseline, and data boundary

How do we decide whether a simple rule or pretrained baseline is enough before training YOLO?

Evidence

A tutor-approved project charter and object-only data plan

InputData in
The lesson question and a fixed non-personal example: How do we decide whether a simple rule or pretrained baseline is enough before training YOLO?
MemoryStored state
The rules and intermediate record kept for the next step: A project charter covering goal, user, misuse, non-AI baseline, success metric, and stop condition
ProcessTransform
Compare, build, and test with these ideas: problem formulation, baseline, data governance, go-no-go.
OutputResult
A tutor-approved project charter and object-only data plan
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

05Design the class ontology and ambiguity policy

How can annotation rules expose rather than hide ambiguous boundary cases?

Evidence

A two-annotator pilot and disagreement log

InputData in
The lesson question and a fixed non-personal example: How can annotation rules expose rather than hide ambiguous boundary cases?
MemoryStored state
The rules and intermediate record kept for the next step: An annotation guide defining classes, inclusion, exclusion, ambiguity, and unknown handling
ProcessTransform
Compare, build, and test with these ideas: ontology, mutual exclusivity, ambiguity, unknown class.
OutputResult
A two-annotator pilot and disagreement log
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

06Validate YOLO box coordinates and annotation errors

How do incorrect normalized center coordinates break boxes after resizing?

Evidence

A report of invalid coordinates, missing labels, extra labels, and class mismatches

InputData in
The lesson question and a fixed non-personal example: How do incorrect normalized center coordinates break boxes after resizing?
MemoryStored state
The rules and intermediate record kept for the next step: An audit script that redraws label text files over images to verify range and class
ProcessTransform
Compare, build, and test with these ideas: xywh, normalization, class index, annotation invariant.
OutputResult
A report of invalid coordinates, missing labels, extra labels, and class mismatches
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

07Prevent leakage with grouped splits

How should object instances and adjacent frames be grouped for an honest generalization test?

Evidence

A split manifest, seed, class-condition distribution, and leakage check

InputData in
The lesson question and a fixed non-personal example: How should object instances and adjacent frames be grouped for an honest generalization test?
MemoryStored state
The rules and intermediate record kept for the next step: A deterministic split script using capture session and object instance as group keys
ProcessTransform
Compare, build, and test with these ideas: group key, near duplicate, distribution shift, held-out set.
OutputResult
A split manifest, seed, class-condition distribution, and leakage check
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

08Audit the dataset separately from augmentation

What is lost when real capture diversity and synthetic augmentation are treated as equivalent evidence?

Evidence

A dataset gap list and an approved augmentation hypothesis

InputData in
The lesson question and a fixed non-personal example: What is lost when real capture diversity and synthetic augmentation are treated as equivalent evidence?
MemoryStored state
The rules and intermediate record kept for the next step: An audit separating class, background, light, scale, occlusion coverage from augmentation plans
ProcessTransform
Compare, build, and test with these ideas: coverage, imbalance, augmentation, domain gap.
OutputResult
A dataset gap list and an approved augmentation hypothesis
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

09Pre-register the transfer learning experiment

How can model, epochs, image size, metrics, and comparisons be fixed before seeing results?

Evidence

A timestamped experiment plan written before execution

InputData in
The lesson question and a fixed non-personal example: How can model, epochs, image size, metrics, and comparisons be fixed before seeing results?
MemoryStored state
The rules and intermediate record kept for the next step: An experiment plan fixing model, data, seed, epochs, image size, and metrics for a baseline and one follow-up run
ProcessTransform
Compare, build, and test with these ideas: transfer learning, pre-registration, hyperparameter, checkpoint.
OutputResult
A timestamped experiment plan written before execution
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

10Track training runs and artifacts

Why preserve configuration, logs, plots, and code commits instead of only best.pt?

Evidence

A baseline training artifact pack and rerun verification

InputData in
The lesson question and a fixed non-personal example: Why preserve configuration, logs, plots, and code commits instead of only best.pt?
MemoryStored state
The rules and intermediate record kept for the next step: An experiment ledger linking run name, source commit, data hash, arguments, and output hash
ProcessTransform
Compare, build, and test with these ideas: run directory, configuration, checkpoint, provenance.
OutputResult
A baseline training artifact pack and rerun verification
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

11Interpret mAP and PR curves through real errors

Why can a model with strong mAP50-95 still fail on one class or small objects?

Evidence

An evaluation report defining metrics, threshold choice, and representative FP, FN, and localization errors

InputData in
The lesson question and a fixed non-personal example: Why can a model with strong mAP50-95 still fail on one class or small objects?
MemoryStored state
The rules and intermediate record kept for the next step: An evaluation notebook linking global, per-class, size, and condition-slice metrics
ProcessTransform
Compare, build, and test with these ideas: IoU, precision-recall curve, mAP50-95, confusion matrix, per-class metric.
OutputResult
An evaluation report defining metrics, threshold choice, and representative FP, FN, and localization errors
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

12Find failure conditions with error slices

How can failures by light, occlusion, scale, and background be exposed behind an average score?

Evidence

Prioritized failure slices and counterexample images

InputData in
The lesson question and a fixed non-personal example: How can failures by light, occlusion, scale, and background be exposed behind an average score?
MemoryStored state
The rules and intermediate record kept for the next step: A failure browser grouping errors by class, size, light, occlusion, background, and cause hypothesis
ProcessTransform
Compare, build, and test with these ideas: error taxonomy, slice, false positive, false negative, localization error.
OutputResult
Prioritized failure slices and counterexample images
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

13Use an ablation to test one claim

How can we claim that one change in data, image size, or augmentation repaired a weak slice?

Evidence

An ablation table covering metric, slice, latency, side effects, and conclusion

InputData in
The lesson question and a fixed non-personal example: How can we claim that one change in data, image size, or augmentation repaired a weak slice?
MemoryStored state
The rules and intermediate record kept for the next step: An ablation comparing the baseline with a one-variable run under the same test and latency conditions
ProcessTransform
Compare, build, and test with these ideas: ablation, control, seed, effect, trade-off.
OutputResult
An ablation table covering metric, slice, latency, side effects, and conclusion
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

14Design a webcam inference pipeline and latency budget

How can capture, preprocessing, inference, postprocessing, and display bottlenecks be measured separately?

Evidence

A latency breakdown, threshold rationale, frame policy, and live demo

InputData in
The lesson question and a fixed non-personal example: How can capture, preprocessing, inference, postprocessing, and display bottlenecks be measured separately?
MemoryStored state
The rules and intermediate record kept for the next step: A Python webcam application with stage timing and a stale-frame policy in an object-only zone
ProcessTransform
Compare, build, and test with these ideas: frame pipeline, latency budget, frame drop, stream buffer, threshold.
OutputResult
A latency breakdown, threshold rationale, frame policy, and live demo
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

15Review robustness, privacy, and licensing before release

What failures and rights must be reviewed before publishing or integrating a working demo?

Evidence

A robustness report, privacy boundary, model card, and release go-no-go decision

InputData in
The lesson question and a fixed non-personal example: What failures and rights must be reviewed before publishing or integrating a working demo?
MemoryStored state
The rules and intermediate record kept for the next step: A blur, glare, occlusion, unknown-object test plus a data, model, and code license checklist
ProcessTransform
Compare, build, and test with these ideas: robustness, failure injection, privacy, model card, license review.
OutputResult
A robustness report, privacy boundary, model card, and release go-no-go decision
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

16Technical defense and a future robot interface contract

What keeps webcam detections from becoming immediate movement when a future Tello or RoboMaster adapter is added?

Evidence

A reproducible repository, live demo, poster, model card, limitations defense, and robot-free interface test

InputData in
The lesson question and a fixed non-personal example: What keeps webcam detections from becoming immediate movement when a future Tello or RoboMaster adapter is added?
MemoryStored state
The rules and intermediate record kept for the next step: A future adapter contract defining detection events, confidence, timestamps, no-detection state, and a separate safety controller
ProcessTransform
Compare, build, and test with these ideas: interface contract, decision gate, watchdog, E-stop, technical defense.
OutputResult
A reproducible repository, live demo, poster, model card, limitations defense, and robot-free interface test
ControlCheck and adjust
Inspect one success and one failure, change one input, stored rule, or process step, and test again.

READ THE REPOSITORY

Read the YOLO GitHub as more than a code folder

A GitHub repository holds more than executable code. Connecting its README, docs, source, issues, releases, and license lets a student explain where a model came from and how to reproduce it.

GitHub · ultralytics/ultralytics ↗

Reviewed baseline763f47e0e5b6Ultralytics 8.4.123yolo26n.pt

ultralytics / ultralyticspublic

README
What it does and how to begin
docs
Official usage and current syntax
ultralytics/
Source implementing models and tasks
issues · releases
Known problems and change history
LICENSE
Terms to review before reuse or deployment

A student drawing bounding boxes around objects

LABEL · SPLIT · TRAIN · CHECK

Train on a small dataset and preserve the evidence

Image count comes after clear class rules, varied scenes, consistent boxes, and a leakage-resistant split. Students record the command, package version, model, data YAML, and result folder together.

pip install ultralytics==8.4.123
yolo detect train model=yolo26n.pt \
  data=data.yaml epochs=40 imgsz=640

Commands were checked against official documentation on 2026-08-20. Current docs are rechecked before class.

PERCEPTION BEFORE MOTION

Webcam now, moving robots later

Webcam detection is perception: it draws boxes on a screen. A future Tello or RoboMaster extension places a separate decision gate and safety controller between a detection and any motion.

This coursewebcam → YOLO → boxes

Future extensiondetection event → safety gate → robot adapter

Object-only webcam detection — Tello · RoboMaster: after a separate safety review

PRIVACY BY DESIGN

Students can learn without filming people

The course uses student-chosen objects, fixed fixtures, and an object-only camera zone. Faces, names, home backgrounds, and private messages stay out of the dataset.

Students control camera on and off
Non-personal objects instead of faces
Separate checks for dataset, model, and code licenses

QUESTIONS

Frequently asked questions

Do elementary students train YOLO themselves?

Students choose the images and box rules while the instructor runs a bounded training job. Students predict outcomes and compare failure cases.

Is a GPU required?

No. The course uses a small nano model and a bounded dataset. A verified output fixture preserves the same analysis when hardware is slow.

Do students just copy GitHub code?

They first reproduce an official quickstart while recording versions, sources, licenses, and changes. Then they vary one condition at a time.

Will students move a Tello or RoboMaster right away?

No. This course stops at webcam perception. A future robot connection is designed with mock events and a safety interface before any separate motion course.

Is the course online only?

No. We run in-person classes and instructor-led online classes. One-to-one and small group formats are set by the student's readiness and schedule.

START WITH THE RIGHT DEPTH

Choose the right starting depth together

Tell us the student's grade, coding experience, and the demo they want to build. We will suggest the right entry point.

Ask about the course