Pseudo-Siamese Blind-spot Transformers for Self-Supervised Real-World Denoising

Authors: Yuhui Quan, Tianxiang Zheng, Hui Ji

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experiments
Researcher Affiliation Academia Yuhui Quan , Tianxiang Zheng School of Computer Science and Engineering South China University of Technology Hui Ji Department of Mathematics National University of Singapore
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No Our work is implemented on Py Torch1.10 and CUDA 11.8, which will be released upon paper acceptance.
Open Datasets Yes Three widely-used real-world datasets are used for evaluation: SIDD [13], DND [14], and NIND [64].
Dataset Splits Yes The SIDD-Medium subset is chosen as training data, consisting of 320 noisy/clean image pairs. The validation subset, denoted by SIDD-Validation, consists of 1280 paired samples for hyper-parameter tuning and ablation study.
Hardware Specification Yes All experiments are conducted on an NVIDIA A6000 GPU.
Software Dependencies Yes Our work is implemented on Py Torch1.10 and CUDA 11.8
Experiment Setup Yes The grid size of Self Former-D is set to image size divided by 8, and it doubles for Self Former-F. ... Self Former-D is optimized using Adam with a learning rate of 0.0001, and that of Self Former-F is doubled. Other parameters of Adam are set to default. The entire model is trained for 30 epochs for full convergence.