Computer Vision in 2 Minutes Each
How machines learn to see. For the full story, read the computer vision series.
Camera model
A camera projects the 3D world onto a 2D image.
Remember: intrinsics describe the lens (focal length, center); extrinsics describe where the camera sits. Needed whenever images meet 3D.
Go deeper: Module 3: Calibration.
Convolution
Slide a small filter across the image to detect a pattern everywhere.
Picture it: a stencil for “vertical edge” pressed against every spot on the photo.
Remember: early layers learn edges and colors; deeper layers learn textures, parts and objects.
CNN architectures
| Model | Remember it as |
|---|---|
| AlexNet (2012) | Deep learning’s ImageNet breakthrough on GPUs |
| VGG | Just stack small 3×3 filters |
| ResNet | Skip connections make 100+ layers trainable |
| MobileNet / EfficientNet | Cheap convolutions for phones |
| ConvNeXt | A modernized CNN that matches transformers |
Vision transformer (ViT)
Cut the image into patches, treat them like words, run a transformer.
Remember: sees the whole image at once; needs lots of data or pretraining, then scales very well.
Data augmentation
Create variations of training images (flips, crops, color changes) so the model learns what doesn’t matter.
Remember: the cheapest big win in vision; augmentations must preserve the label.
Object detection
Find each object and draw a box around it with a label.
Remember: two-stage detectors (Faster R-CNN) are accurate; one-stage (YOLO) are fast; DETR treats detection as predicting a set of objects with a transformer.
IoU and NMS
IoU measures box overlap; NMS removes duplicate boxes for the same object.
Remember: mAP, the main detection metric, counts a box as correct above an IoU threshold.
Segmentation
| Type | Output |
|---|---|
| Semantic | A class for every pixel (“road,” “car”) |
| Instance | A separate mask for each object |
| Panoptic | Both at once |
Remember: U-Net’s encoder-decoder with skip connections is the classic architecture.
Segment Anything (SAM)
A foundation model that segments whatever you point at, click or box.
Remember: promptable segmentation without task-specific training.
Self-supervised vision
Learn visual features without labels.
Remember: masked autoencoders hide patches and rebuild them; DINO learns features so good that object boundaries appear without supervision.
Depth and 3D
Recover distance from stereo cameras, motion, or a single image with a learned model.
Remember: monocular depth models (Depth Anything) give relative depth; a few known distances fix the scale. Point clouds come from lidar or depth.
NeRF and Gaussian splatting
Reconstruct a 3D scene from photos so you can render new viewpoints.
Remember: NeRF stores the scene in a neural network (slow to render); Gaussian splatting uses millions of small blobs (renders in real time).
Bird’s-eye view
Turn camera views into a top-down map around a vehicle.
Go deeper: Module 6: Perception.
CLIP
Images and captions trained to land close together in one space.
Remember: enables zero-shot classification (“a photo of a dog” vs “a photo of a cat”) and powers text-to-image models.
Vision-language models
A vision encoder plus an LLM that can talk about images.
Remember: vision encoder → small connector → language model.
Go deeper: Qwen3-VL, When AI Sees and Speaks.
Generating images and video
Go deeper: Generative Models cheatsheet.