Binarized Spectral Compressive Imaging

Authors: Yuanhao Cai, Yuxin Zheng, Jing Lin, Xin Yuan, Yulun Zhang, Haoqian Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 ExperimentComprehensive quantitative and qualitative experiments manifest that our proposed Bi SRNet outperforms state-of-the-art binarization algorithms.
Researcher Affiliation Academia Yuanhao Cai 1, Yuxin Zheng 1, Jing Lin 1, Xin Yuan 2, Yulun Zhang 3, , Haoqian Wang 1, 1 Tsinghua University, 2 Westlake University, 3 ETH Zürich
Pseudocode No The paper provides architectural diagrams (e.g., Fig. 2) but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Code and models are publicly available at https://github.com/caiyuanhao1998/Bi SCI
Open Datasets Yes Two simulation datasets, CAVE [66] and KAIST [65], are adopted. The CAVE dataset provides 32 HSIs with a spatial size of 512 512. The KAIST dataset includes 30 HSIs at a spatial size of 2704 3376. We use CAVE for training and select 10 scenes from KAIST for testing.
Dataset Splits No The paper mentions using CAVE for training and KAIST for testing, and describes how training samples are generated (patches, data augmentation). However, it does not explicitly state a separate validation set or provide details on validation splits.
Hardware Specification Yes We train Bi SRNet for 300 epochs on a single RTX 2080 GPU.
Software Dependencies No The paper states, 'The proposed Bi SRNet is implemented by Py Torch [67]', but it does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes We use Adam [68] optimizer (β1 = 0.9 and β2 = 0.999) and Cosine Annealing [69] scheduler to train Bi SRNet for 300 epochs on a single RTX 2080 GPU. Training samples are patches with spatial sizes of 256 256 and 96 96 randomly cropped from 28-channel 3D HSI data cubes for simulation and real experiments. The shifting step d is 2. The batch size is 2. We set the basic channel C = Nλ = 28 to store HSI information. We use random flipping and rotation for data augmentation. The training loss function is the root mean square error (RMSE) between reconstructed and ground-truth HSIs.