Learning Real-World Image De-weathering with Imperfect Supervision
Authors: Xiaohui Liu, Zhilu Zhang, Xiaohe Wu, Chaoyu Feng, Xiaotao Wang, Lei Lei, Wangmeng Zuo
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on two real-world de-weathering datasets show that our method helps existing de-weathering models achieve better performance. Code is available at https://github.com/1180300419/ imperfect-deweathering. Experiments are conducted with Rain Robust (Ba et al. 2022) and Restormer (Zamir et al. 2022) models on GT-Rain-Snow (Ba et al. 2022) and Weather Stream (Zhang et al. 2023) datasets. With the proposed method, the quantitative and qualitative results of existing de-weathering models are greatly improved, while not increasing any inference cost. |
| Researcher Affiliation | Collaboration | Xiaohui Liu1, Zhilu Zhang1, Xiaohe Wu1*, Chaoyu Feng2, Xiaotao Wang2, Lei Lei2, Wangmeng Zuo1 1Harbin Institute of Technology 2Independent Researcher |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/1180300419/ imperfect-deweathering. |
| Open Datasets | Yes | Experiments are conducted with Rain Robust (Ba et al. 2022) and Restormer (Zamir et al. 2022) models on GT-Rain-Snow (Ba et al. 2022) and Weather Stream (Zhang et al. 2023) datasets. |
| Dataset Splits | No | The paper states using 'two training datasets' and 'the testing dataset proposed by (Zhang et al. 2023)', but does not provide specific training/validation/test split percentages or sample counts. |
| Hardware Specification | No | Experiments based on Rain Robust are conducted on a single GPU. The corresponding patch size is 256 256, and the batch size is 8. Experiments based on Restormer are conducted on two GPUs. The corresponding patch and batch sizes are set to 168 168 and 6, respectively. All experiments are implemented with the Py Torch framework. (No specific GPU model, CPU, or other hardware details are provided.) |
| Software Dependencies | No | All experiments are implemented with the Py Torch framework. (No specific version number for PyTorch or other libraries is provided.) |
| Experiment Setup | Yes | For optimizing the models, we adopt the Adam with β1 = 0.9 and β2 = 0.999. A warm-up strategy is employed to gradually increase the learning rate from 5 10 5 to 2 10 4, followed by the cosine annealing strategy to decrease the learning rate from 2 10 4 to 10 6. Experiments based on Rain Robust are conducted on a single GPU. The corresponding patch size is 256 256, and the batch size is 8. Experiments based on Restormer are conducted on two GPUs. The corresponding patch and batch sizes are set to 168 168 and 6, respectively. |