Semi-Supervised Learning

• Semi-supervised Learning

뮌헨공대 (Technical University of Munich)에서 공부한
[IN2375 Computer Vision - Detection, Segmentation and Tracking] 컴퓨터비전 노트 정리

Semi-supervised Learning

아이디어

use both labelled and unlabelled data

assumption

• smoothness : if two inputs are close, their labels are same
• low density : decision boundary should pass through region with low density
• manifold : data come from multiple low-dim. manifolds if data points share same manifold, their labels are same

unsupervised pre-processing

feature extraction

wrapper method 중 self-training

OnAVOS는 slow이므로 first frame/new frame 대신
offline으로 labelled/unlabelled data 사용

initial prediction이 중요하므로 미리 train strong baseline on labelled set

energy minimization (low-density assumption 적용)

minimize \(-p(x_i)logp(x_i)\) = entropy of class distribution of each pixel \(x_i\)

VAN (virtual adversarial network) (smoothness assumption 적용)

labelled set : true posterior(gt) 와 adversarial 추가한 image의 prediction 비교

unlabelled set : 기존 image의 prediction 과 adversarial 추가한 image의 prediction 비교

Domain Alignment

GAN의 원리 사용하여 unlabelled real data와 labelled synthetic data의 distribution을 비슷하게

Consistency Regularization

image에 transformation을 가하더라도 robust하게 consistent prediction을 하도록 consistency loss 추가