10.6.1 Project Roadmap: Build a Vision Evidence Pack

A computer vision project is not “I used a model.” It is a loop of data, annotation, model output, metrics, failure cases, and presentation.

See the Project Loop First

Start with classification if you need the fastest complete loop. Move to detection for boxes, segmentation for masks, and OCR/video/3D for specialized scenarios.

Run a Project Readiness Check

Use this before you call the project presentable.

project = {
    "task": "helmet detection",
    "has_data_note": True,
    "has_metric": True,
    "has_failure_case": True,
    "has_annotation_rule": True,
}

ready = all(project[key] for key in ["has_data_note", "has_metric", "has_failure_case", "has_annotation_rule"])

print("task:", project["task"])
print("presentable:", ready)

Expected output:

task: helmet detection
presentable: True

If a project has no annotation rule or failure case, it is still a demo, not a portfolio project.

Learn in This Order

Step	Project Type	Evidence
1	Classification	Dataset split, accuracy/F1, confusion examples
2	Detection	Box annotations, IoU/mAP, false positives and missed detections
3	Segmentation	Masks, IoU/Dice, boundary failures
4	Industry scenario	Risk notes, user impact, deployment idea
5	Hands-on workshop	Reproducible mini pipeline before larger project pages

Run 10.6.4 Hands-on: Build a Reproducible Vision Mini Pipeline before expanding the project.

Project Deliverable Standards

Deliverable	Minimum Requirement	Stronger Portfolio Version
README	Goal, run command, dependencies, examples	Add task boundary, data source, deployment idea
Data and annotation	Image source, class list, annotation format	Add annotation examples, quality checks, bias notes
Results	At least 1 input image and prediction result	Add correct, false positive, false negative, boundary cases
Evaluation	Accuracy, F1, mAP, IoU, Dice, or OCR hit rate	Add error analysis by class, scenario, lighting, clarity
Failure analysis	At least 1 real failure	Add suspected cause, fix action, regression check
Presentation	Screenshot or short GIF proving it runs	Build a clear visual project page

Pass Check

You pass this chapter when your vision project can be reproduced, has clear data and annotation rules, reports proper metrics, and shows where the model fails.