MambaLLIE: Implicit Retinex-Aware Low Light Enhancement with Global-then-Local State Space

Authors: Jiangwei Weng, Zhiqiang Yan, Ying Tai, Jianjun Qian, Jian Yang, Jun Li

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that Mamba LLIE significantly outperforms state-of-the-art CNN and Transformer-based methods.
Researcher Affiliation Academia 1School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, 210094, China 2School of Intelligence Science and Technology, Nanjing University, Suzhou, 215163, China
Pseudocode No The paper provides architectural diagrams (Figure 2) and describes modules in text but does not include any formal pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at Project Page.
Open Datasets Yes We employ five paired low light image datasets for evaluation, including LOL-V2-real [59], LOL-V2-syn [59], SMID [4], SDSD-indoor [48] and SDSD-outdoor [48] datasets.
Dataset Splits Yes LOL-V2-real contains 689 low-normal light paired images for training and 100 pairs for testing; LOL-V2-syn includes 900 paired images for training and the 100 pairs for testing; Besides, SMID is composed of the 15763 short-long exposure paired images for training and the remaining images for testing; SDSD-indoor and SDSD-outdoor are all subsets of SDSD dataset (the static version), which extract the paired images from 62 and 116 pairs for training, and the left 6 and 10 pairs for testing.
Hardware Specification Yes We implement Mamba LLIE in Py Torch [40] on a server with the 4090GPUs.
Software Dependencies No The paper mentions 'Py Torch [40]' but does not specify a version number for it or any other software dependencies.
Experiment Setup Yes Random cropping the image pairs into 128 x 128 patches as training samples, data augmentation is performed on the training samples such as rotation and flipping. The batch size is 8. In terms of optimization procedure, Adam [24] is adopted as the optimizer with β1 = 0.9 and β2 = 0.999; The training iterations is set to 1.5 x 10^5. The initial e learning rate is set to 2 x 10^-4 and steadily decreased by by the cosine annealing scheme. The loss criterion is mean absolute error (MAE), thus peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) [52] is selected as the evaluation metrics for the paired datasets.