Networks are Slacking Off: Understanding Generalization Problem in Image Deraining

Authors: Jinjin Gu, Xianzheng Ma, Xiangtao Kong, Yu Qiao, Chao Dong

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through comprehensive and systematic experimentation, we discover that this strategy does not enhance the generalization capability of these networks. Our experiments reveal that better generalization in a deraining network can be achieved by simplifying the complexity of the training background images.
Researcher Affiliation Academia Jinjin Gu1,2 Xianzheng Ma1 Xiangtao Kong1,3 Yu Qiao1,3 Chao Dong1,3, 1 Shanghai AI Laboratory 2 The University of Sydney 3 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences jinjin.gu@sydney.edu.au, xianzhengma@pjlab.org.cn {xt.kong, yu.qiao, chao.dong}@siat.ac.cn
Pseudocode No The paper describes methods and processes but does not include any explicitly labeled "Pseudocode" or "Algorithm" blocks.
Open Source Code No The paper states: "All the trained models and code will be publicly available." This is a promise for future release, not concrete availability at the time of publication.
Open Datasets Yes We select image distribution as the second aspect of our dataset construction. We sample from four image datasets that are distinct from each other: Celeb A (face images) [35], DIV2K (natural images) [50], Manga109 (comic images) [38], and Urban100 (building images) [24].
Dataset Splits No The paper describes training datasets and test sets, but it does not explicitly mention or specify a separate validation dataset split with details like percentages or sample counts.
Hardware Specification Yes All models are built using the Py Torch framework [39] and trained with NVIDIA A100 GPUs.
Software Dependencies No The paper mentions "Py Torch framework [39]" and "Adam for training" but does not provide specific version numbers for these or any other software components necessary for replication.
Experiment Setup Yes We use Adam for training. The initial learning rate is 2 10 4 and β1 = 0.9, β2 = 0.99. For each network, we fixed the number of training iterations to 250,000. The batch size is 16. Input rainy images are of size 128 128. The cosine annealing learning strategy is applied to adjust the learning rate. The period of cosine is 250,000 iterations.