-
Notifications
You must be signed in to change notification settings - Fork 0
UPerNet
Kim Na Young edited this page May 2, 2022
·
2 revisions
scene recognition, object detection, texture recognition, material recognition taskλ₯Ό λμμ ν΄κ²°νλ frameworkλ₯Ό μ μ
- FPN
- PPM
- Head
-
FPN (Feature Pyramid Network)
- multi-level feature representationsλ₯Ό μ¬μ©νλ feature extractor
- pyramidal hierarchy ꡬ쑰
- top-down architecture + lateral connections -> high-level semantic informationλ₯Ό middle νΉμ low level informationκ³Ό μ΅ν©
-
PPM (Pyramid Pooling Module)
- backbone networkμ λ§μ§λ§ layerμ μΆκ°
- FPNμ top-down branchλ‘ μ§ννκΈ° μ μ PPM μμΉ
- ν¨κ³Όμ μΈ global prior representations
-
Head
- scene recognition, object detection, texture recognition, material recognition taskλ₯Ό λμμ ν΄κ²° κ°λ₯νλλ‘ λ§λ¦
- single networkμμ, multiple levelμ μ‘΄μ¬νλ visual attributesλ₯Ό parseνκ³ unify ν μ μμ
- Scene Head / Object Head / Part Head / Material Head / Texture Head
- scene recognition, object detection, texture recognition, material recognition taskλ₯Ό λμμ ν΄κ²° κ°λ₯νλλ‘ λ§λ¦
- Conv 3x3 -> GAP -> Classifierλ‘ κ΅¬μ±
- image-levelμ highest-level informationμ΄ νμνλ―λ‘, GAP μ μ©
- Conv 3x3 -> Classifierλ‘ κ΅¬μ±
- μ€ν κ²°κ³Ό, FPNμ λͺ¨λ feature mapμ μ΅ν©νμ¬ μ¬μ©νλ κ²½μ°κ°, highest resolutionμ feature mapλ§ μ¬μ©νλ κ²½μ°λ³΄λ€ κ²°κ³Όκ° λ μ’μμ
- material recognitionμ μν΄μ context informationκ³Ό local featureκ° νμ
- λͺ¨λ feature mapμ μ΅ν©ν΄μ μ¬μ©νλ λμ , highest resolutionμ feature mapμ μ¬μ©
- μ¬λ¬ convolutional layerλ₯Ό μΆκ°νλ κ²μΌλ‘ ꡬμ±
- 맀 pixelλ§λ€ texture labelμ μμΈ‘νλλ‘ λ§λ¦
- Texture branchμ gradientλ back-propagation λμ§ μμ -> Texture Headλ§ νμ΅νλλ‘
Reference
Human Pose Estimation
CNN Visualization
Image Generation
Multi-modal Learning