Robust Feature Rectification of Pretrained Vision Models for Object Recognition

Authors: Shengchao Zhou, Gaofeng Meng, Zhaoxiang Zhang, Richard Yi Da Xu, Shiming Xiang

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Evaluations on CIFAR-10 and Tiny-Image Net demonstrate that the accuracy of ROFER is 5% higher than that of SOTA methods on different degradations.
Researcher Affiliation Academia 1 NLPR, Institute of Automation, Chinese Academy of Sciences, 2 School of Artificial Intelligence, University of Chinese Academy of Sciences, 3 CAIR, HK Institute of Science and Innovation, Chinese Academy of Sciences, 4 FSC1209, Kowloon Tong Campus, Hong Kong Baptist University
Pseudocode No The paper describes the method using text and diagrams (Figure 3), but does not contain a formal 'Pseudocode' or 'Algorithm' block.
Open Source Code No The paper states 'Codes are implemented in Python of 3.6.2 and Pytorch of 1.7.1.' but does not provide a link or explicit statement about the code being open-source or publicly available.
Open Datasets Yes Datasets and Backbones CIFAR-10 combined with CIFAR-100 and Tiny-Image Net combined with CUB-2002011 are selected datasets for evaluation. When testing on CIFAR-10 (or Tiny-Image Net), only CIFAR-100 (or CUB200-2011) is used to train ROFER. Following the generation method in (Hendrycks and Dietterich 2019), every dataset has clear and generated degraded images, where there are five levels of intensity for each degradation.
Dataset Splits No The paper mentions 'The training epoch is 30' and refers to training and testing datasets, but does not explicitly describe a separate validation split or cross-validation methodology for reproduction.
Hardware Specification Yes Training and testing are finished on a single TITAN RTX GPU.
Software Dependencies Yes Codes are implemented in Python of 3.6.2 and Pytorch of 1.7.1.
Experiment Setup Yes The training batch size is 100 and the optimizer is SGD with 0.01 as the learning rate, 0.9 as the momentum. The training epoch is 30 and the scaling parameter α in Eq.(8) is set to be 10.