Computer Vision

Images And Convolution

1. Images as TensorsPixels with shape: height, width, channels, and the tensor contract.1–2 h
2. ConvolutionThe sliding weighted sum: locality, sharing, and one honest hand computation.1–2 h
3. Filters and Feature MapsPatterns become channels: what filters detect and what maps record.1–2 h
4. Stride and PaddingResolution control: how steps and borders set the output size.1–2 h
5. PoolingDownsampling with a purpose: invariance intuition and what pooling discards.1–2 h

Architectures

1. CNN HierarchyEdges to parts to objects: receptive fields and the compositional ladder.1–2 h
2. CNN Architectures and Residual NetworksFrom LeNet to ResNet: depth, skip connections, and the design moves that made deep vision trainable.1–2 h
3. Vision TransformersImages as sequences of patches processed by a transformer, trading convolution's built-in bias for data and scale.1–2 h
4. Self-Supervised Visual RepresentationsLearning visual features without labels: contrastive views, masked image modelling, and self-distillation.1–2 h

Tasks

1. Object DetectionPredicting what and where: boxes, anchors or queries, IoU, and non-maximum suppression.1–2 h
2. Semantic and Instance SegmentationPer-pixel prediction: encoder-decoder networks, upsampling, and separating objects of the same class.1–2 h
3. Video UnderstandingAdding time: frame sampling, 3D and factorised convolutions, and temporal attention.1–2 h

Geometry

1. 3D Vision: Cameras, Depth, and Scene RepresentationsFrom pixels to 3D: the pinhole camera, stereo depth, and learned scene representations such as NeRF.1–2 h