Computer vision for the foundation-model era: detection, try-on, inspection and generation.

Images as data, and how vision models learn from them.

1. Pixels, channels and transforms
2. Convolutions and feature maps
3. From CNNs to vision transformers
4. The main CV tasks
5. Classic OpenCV still worth knowing

Train accurate classifiers with little data.

1. Pretrained backbones
2. Fine-tuning strategies
3. Data augmentation
4. Class imbalance
5. Evaluating classifiers

Detect and track objects with the YOLO family.

1. How one-stage detectors work
2. Training YOLO on custom data
3. NMS-free detectors
4. Tracking across video frames
5. mAP and speed metrics

Pixel-accurate masks from points, boxes or text prompts.

1. Semantic vs instance segmentation
2. Promptable segmentation
3. Concept prompts with SAM 3
4. Video segmentation and tracking
5. Using SAM to auto-label data

Find objects you never trained on.

1. Text-prompted detectors
2. YOLOE and Grounding DINO
3. Accuracy vs edge speed
4. When to fine-tune instead
5. Combining closed and open models

Models that read images and answer questions about them.

1. How VLMs connect vision and text
2. Visual question answering
3. Document and OCR understanding
4. Grounding answers in regions
5. Evaluating VLM output

The work that decides most of your accuracy.

1. Labelling workflows
2. Auto-labelling with foundation models
3. Dataset quality checks
4. Synthetic data
5. Versioning datasets

Generate and edit images: try-on, inpainting and more.

1. Diffusion models in brief
2. Image editing and inpainting
3. Virtual try-on systems
4. Controllable generation
5. Ethics and misuse

Ship vision models that are fast and cheap enough.

1. Export to ONNX and TensorRT
2. Quantisation for edge devices
3. Video pipelines at scale
4. Serving APIs
5. Monitoring in production

Build, evaluate and deploy a complete system.

1. Picking a real use case
2. Data to deployment
3. Measuring real-world accuracy
4. Demo and write-up
5. Portfolio presentation