[Paper] Self-Supervised Visual Feature Learning With Deep Neural Networks: A Survey

“Self-Supervised Visual Feature Learning With Deep Neural Networks: A Survey”이란 논문에 대한 리뷰입니다.

원문은 링크에서 확인할 수 있습니다.

Insight

• First train general attributes
• Shallower blocks focus on the low-level general features

• Deeper blocks focus on the high-level task specific features

• Effective pretext tasks ensure that semantic features learned through the process of accomplishing the pretext tasks.

Pseudo labels without human annotations

• Minimizing error between the prediction & pseudo labels

Schema

Pre-training with pretext

• (Unlabelled, Self-supervised) Generation based
• Context based
• Free semantic label based
• Cross modal based

Downstream Task (Labelled) Image Classification

• Sematic segmentation
• Object detection
• Human action recognition

Feature Learnings description

• Generation based Image Generation GAN
• Video Generation
• Context based Context Similarity Between image patches
  • Spatial Context Similarity Spatial relation among image patches
  • Temporal Context Similarity Frame sequence (Video)
• Free semantic label based
• Cross modal based Two different channels of input data correspond to each other

Downstream Task Describe

• Image Classification (Quality of feature)
• Sematic segmentation (Generality of feature) -Assign semantic labels to each pixel (Pixelwise annotation)
• Object detection (Generality of feature)
• Localizing the position of objects
• Proposals based on feature map
• FC to bound box of object
• Human action recognition (Quality of feature)

Feature Learning

• Generation based
  • Pseudo labels are images themselves
• Context based
  • (Common) clustering
  • Predictive task for group ID
  • Contrastive task for distance btw features
• Free semantic label based
• Cross modal based

Qualitative evaluation

• Kernel Visualization Compare first convolution layer with supervised one
• Feature Map Visualization Attention of networks