Career roadmap
Computer Vision Engineer
Make machines see: detection, segmentation and recognition, running fast enough to be useful.
Before you start Computer Vision
- Python and NumPy
- Linear algebra and calculus basics
- Neural network fundamentals
Image fundamentals
Classical vision still solves many problems faster and cheaper than a network.
Colour spaces and sampling explain a surprising number of production bugs.
- Colour spaces and conversions
- Resolution, sampling and aliasing
- Image formats and compression artefacts
- Camera parameters and exposure
Filters, edges and morphology — often sufficient and always fast.
- Convolution, filtering and denoising
- Edge and corner detection
- Morphological operations
- Thresholding and contours
Essential for any system where measurements matter.
- Camera models and intrinsics
- Distortion and calibration
- Homography and perspective transforms
- Stereo and depth basics
Still used for tracking, stitching and registration.
- Keypoint detectors and descriptors
- Feature matching and RANSAC
- Image registration
- Optical flow
Vision projects live or die on annotation quality and consistency.
- Annotation formats and tooling
- Labelling guidelines and agreement
- Class imbalance in detection
- Train/validation splits that avoid leakage
BuildBuild a classical pipeline that solves a real inspection task without any deep learning.
Deep learning for vision
The core model families and how to train them without fooling yourself.
Still dominant for many production vision tasks, especially at the edge.
- Convolution, pooling and receptive fields
- Backbone architectures and scaling
- Batch normalisation and regularisation
- Transfer learning from pretrained backbones
The current state of the art at scale, with different data requirements.
- Patch embeddings and attention for images
- Data efficiency versus CNNs
- Hybrid architectures
- Self-supervised pretraining
The most commercially requested vision task.
- One-stage versus two-stage detectors
- Anchor-based and anchor-free approaches
- Non-maximum suppression
- mAP and IoU evaluation
Where precise boundaries matter: medical, industrial and mapping.
- Semantic versus instance segmentation
- Encoder-decoder architectures
- Promptable segmentation models
- Dice, IoU and boundary metrics
Augmentation and validation discipline decide whether results transfer.
- Augmentation strategies that reflect reality
- Learning rate schedules and warmup
- Handling small datasets
- Diagnosing overfitting versus underfitting
BuildTrain a detector on a custom dataset and report mAP with a proper error breakdown.
Applied vision systems
Real deployments have cameras, lighting, motion and awkward edge cases.
Temporal consistency changes both the architecture and the evaluation.
- Multi-object tracking
- Temporal smoothing and re-identification
- Frame sampling strategies
- Action recognition basics
The fastest-moving area, and increasingly the fastest route to a prototype.
- CLIP-style contrastive models
- Zero-shot classification and retrieval
- Multimodal LLMs for visual questions
- When a VLM beats a trained detector
Robotics, AR and industrial measurement need geometry, not just pixels.
- Depth estimation approaches
- Point clouds and processing
- Neural rendering basics
- Sensor fusion
The model works in the lab and fails on the factory floor. This is the classic failure.
- Lighting and camera variation
- Domain adaptation techniques
- Synthetic data and sim-to-real
- Continuous data collection
Most industrial vision runs on constrained hardware, not a cloud GPU.
- Quantisation and pruning
- TensorRT, OpenVINO and mobile runtimes
- Latency and power budgets
- Hardware selection
BuildDeploy a vision system on real captured footage and document every failure mode you found.
Production and MLOps
Vision models degrade quietly as the world changes in front of the camera.
Throughput engineering for image and video workloads.
- Batching and GPU utilisation
- Streaming video ingestion
- Preprocessing on GPU
- Scaling for burst load
A camera moved, the lighting changed, and accuracy fell without any alert.
- Input distribution monitoring
- Confidence distribution shifts
- Sampling outputs for review
- Automated retraining triggers
Reproducibility matters more in vision because datasets are large and messy.
- Dataset versioning
- Experiment tracking
- Reproducible training runs
- Model registry and rollback
Ongoing annotation is a cost centre you will be asked to optimise.
- Active learning for label efficiency
- Quality control for annotators
- Pre-labelling with the current model
- Cost per labelled example
Face and biometric applications carry legal constraints that vary sharply by region.
- Biometric regulation and consent
- Bias across demographic groups
- Surveillance and proportionality
- Documentation and model cards
BuildShip a monitored vision service with drift detection and a labelled feedback loop.
Interview preparation
Interviews cover classical vision, deep learning and deployment constraints.
Convolution arithmetic and detection metrics are asked directly.
- Receptive field and output size calculation
- IoU, mAP and their pitfalls
- NMS behaviour and alternatives
- Augmentation choices and reasons
Design a vision system for a described physical environment.
- Camera placement and constraints
- Model choice for the latency budget
- Handling occlusion and lighting
- Data collection strategy
Image manipulation and pipeline code, sometimes with NumPy only.
- Implementing IoU and NMS from scratch
- Image preprocessing pipelines
- PyTorch dataset and training loop
- Debugging a training run
Accuracy dropped in production. Diagnose it.
- Distinguishing data drift from model bug
- Deciding between more data and a better model
- Edge deployment trade-offs
- Cost per inference reasoning
Real-world data beats benchmark reproductions.
- A project on data you captured
- Metrics with error analysis
- An edge deployment with latency numbers
- Documented failure modes
BuildA public project on real captured data with metrics and a failure analysis.
Computer Vision tools on your CV
- PyTorch
- OpenCV
- Ultralytics / detection frameworks
- TensorRT / OpenVINO
- Albumentations
- Weights & Biases
- CVAT
What Computer Vision employers ask to see
- A detector trained on data you collected, with mAP reported
- An edge deployment with measured latency and power
- A documented domain-gap investigation
- An active learning loop that reduced labelling cost
Manufacturing quality control, medical imaging, retail analytics, autonomous systems and robotics. Less crowded than LLM work, with strong demand in industrial settings.
Content last reviewed 2026-08-31. Guidance only — no institute or paid placement is endorsed anywhere in this book.