Syllabus

Logistic, Introduction, History

Feature extraction and classic descriptors; Fitting: Least Square, RANSAC, Hough Voting; Optimization: First-order/Second-order methods.

Supervised Learning; Linear classifier and Logistic Regression; MLP and Backpropagation; CNN;

Why CNN is better? Training Neural Networks: data preprocessing, weight initialization, optimizer, learning rate Improve CNN training: underfit (Batchnorm, ResNet)

CNN training: overfit; classification, segmentation

Improve CNN training: overfit; classification, segmentation; Transformations; Camera Model: orthographic, weakly persepctive, perspective

Camera calibration; Epipolar Geometry: Vanishing lines/points, Essential matrix, Fundamental matrix; Passive and active stereo system;

Depth sensors; 3D data (voxel, mesh)

3D data (SDF, point cloud); 3D Deep Learning: Point, Mesh; 3D Deep Learning: Sparse Voxel Conv

2D Object detector (SSD, RCNN series, YOLO); Instance Segmentation, Panoptic Segmentation; 3D object detection and instance segmentation

6D pose; rotation representations: Euler angle, axis angle, quaternion; Instance-level 6D pose estimation: PoseCNN, rotation regression; category-level pose estimation: NOCS, orthogonal procrustes; two-frame motion, optical flow.

RNN, LSTM, GRU; Video Analysis: 3D CNN, Two-stream network, ConvRNN

PixelRNN/CNN， VAE

GAN, conditional generative model, learning-based single view reconstruction

Self-supervised Learning and Semi-supervised Learning: pretext tasks, contrastive learning, FixMatch; Neural rendering: neural radiance field (NeRF); Embodied AI.