[SSDA][CLS] MARRS: Modern Backbones Assisted Co-training for Rapid and Robust Semi-Supervised Domain Adaptation

[SSDA][CLS] MARRS: Modern Backbones Assisted Co-training for Rapid and Robust Semi-Supervised Domain Adaptation

• paper : https://openaccess.thecvf.com/content/CVPR2023W/ECV/papers/Jain_MARRS_Modern_Backbones_Assisted_Co-Training_for_Rapid_and_Robust_Semi-Supervised_CVPRW_2023_paper.pdf

• CVPRW 2023 accepted (인용수: 0회, ‘23.10.30기준)

• downsteram task : SSDA for Classification

• Motivation

  • SSDA for classification에 사용되던 baseline 모델 (Res-34)이 out-dated임
  • 단순한 SSL기법 적용한 최신 모델을 활용해서 (CovNext-L) baseline모델 상회

• Contribution

  • SSDA task에 최신 CNN, Vistion-Transformer 모델을 활용한 Co-training 기반 backbone으로 교체 적용

  • Diverse feature 학습을 위해 3가지 모듈 적용 (image-level, feature-level, backbone-level)

    

  • SSDA Datazset SOTA 달성

  • KD 활용한 Mbv2 모델로도 SSDA Dataset SOTA달성

• Overview

  

Feature Extraction

• Backbone level diversity Module (BD)
  • CNN based : ConvNext-XL
  • Vistion-Transformer based : Swin-L
• Image level diversity Module (ID)
  • training 시 weak-aug (psi)를 한쪽 모델에 주입
• Feature distribution level diversity Module (FD)
  • CORAL 을 한쪽 모델 출력에 연결

Classifier Training

• 각 backbone의 confidence prediction을 pseudo-gt로 활용해서 학습 → Co-Training

• Loss

  • Supervised Loss : Labeled source db + Labeled target db (CE Loss)

    

    • n : D dataset에 속한 sample 갯수
    • K : Class 갯수
    • F : classifier F의 출력
    • Label smoothing을 통해 source에 over-confidence 방지

  • Unsupervised Loss

    

    • Unlabeled target db (CELoss) : Co-training
    • Consistency regularization Loss : Strong-aug feature와 original feature간의 일치하도록 강제
      • Strong Aug = RandAugment 사용

• KD

  

  • Inference 속도 향상을 위해 적용
    • Teacher : Swin-L + CovNext-XL (MARRS)
    • Student : MobileNet-V2

• Algorithm

  

• Experiment

  

  

• Ablation Studies

  

  

  

• Inference Time