[Paper Review] Research List-up
🔺 목록 최신화 - # 2024년 2월 1주차
Reading Papers
- [ICML 2020] SHOT : Do we really need to access the source data?
- Source Data에 대한 접근 없이 Source Model만을 사용하는 test-time UDA 기법
- Source Classifier를 고정(freeze)하고, Target Data에 대한 특징 분포가 비슷해지도록 하여, 새로운 도메인에 모델을 적응시킴
- UDA에서 TTA로 넘어가는 개념 초안
- [ICLR 2021] Tent: Fully Test-Time Adaptation by Entropy Minimization
- Entropy Minimization을 사용하여 Test-Time Adaptation을 수행
- Source Data에 대한 접근 없이 Pre-trained model만을 이용하여 Test data를 적응시키는 Fully Test-time Adaptation 기법
- 모델의 배치 정규화 매개변수를 실시간으로 조정하여 Entropy를 줄이고, Confidence를 높혀 일반화 성능을 개선
- [CVPR 2022] CoTTA : Continual Test-Time Adaptation
- 연속적인 Test-time Adaptation을 위한 프레임워크
- 시간이 지나면서 변하는 데이터 분포에 모델을 지속적으로 적응시키기 위해, 예측 Error Accumulation을 최소화하면서도 이전 지식을 잊지않도록 Stochastic Restoration 기법을 사용
- [CVPR 2022] LAME : Parameter-free online testtime adaptation
- [arXiv 2022 Aug 16] GTTA : Gradual Test-Time Adaptation by Self-Training and Style Transfer
- [NeurlPS 2022] NOTE: Robust Continual Test-time Adaptation Against Temporal Correlation
- [CVPR 2023 Mar 22] Robust Mean Teacher for Continual and Gradual Test-Time Adaptation
- By analyzing the gradient properties, we motivate and propose that in the setting of TTA, the symmetric cross-entropy is better suited for a mean teacher than the commonly used cross-entropy.
- cross-entropy 대신 symmetric cross-entropy를 쓰면 성능 개선이 이루어짐을 밝힘.
- [arXiv 2023 Mar 12] Evaluating Continual Test-Time Adaptation for Contextual and Semantic Domain Shifts
- [CVPR 2023 March 23] RoTTA : Robust Test-Time Adaptation in Dynamic Scenarios
- Introduce a new test-time adaptation setup that aligns better with real-world applications, namely Practical Test-Time Adaptation (PTTA).
- PTTA con- siders both continual covariate shift and continual label shift, making it more practical and challenging.
- [WACV 2023 Sep' 23] TeST: Test-time Self-Training under Distribution Shift
- [CVPR 2023 Oct 7] GRoTTA : Generalized Robust Test-Time Adaptation in Dynamic Scenarios
- RoTTA의 후속 논문으로 꽤 많이 개선되어서 나옴
- [ICLR 2023] DELTA: DEGRADATION-FREE FULLY TEST-TIME ADAPTATION*
- [arXiv 2023] Benchmarking Test-Time Adaptation against Distribution Shifts in Image Classification
- [WACV 2024 Oct' 22] Universal Test-time Adaptation through Weight Ensembling, Diversity Weighting, and Prior Correction
- [AAAI 2024 Sep' 26] TRIBE : Towards Real-World Test-Time Adaptation: Tri-Net Self-Training with Balanced Normalization