Deeply Learned Spectral Total Variation Decomposition

Authors: Tamara G. Grossmann, Yury Korolev, Guy Gilboa, Carola Schoenlieb

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we describe the experiments we conducted to evaluate the performance of our proposed TVspec NET.
Researcher Affiliation Academia Tamara G. Grossmann DAMTP University of Cambridge Cambridge, UK tg410@cam.ac.uk Yury Korolev DAMTP University of Cambridge Cambridge, UK yk362@cam.ac.uk Guy Gilboa Department of Electrical Engineering Technion Israel Institute of Technology Haifa, Israel guy.gilboa@ee.technion.ac.il Carola-Bibiane Schönlieb DAMTP University of Cambridge Cambridge, UK cbs31@cam.ac.uk
Pseudocode No The paper describes the network architecture and the loss function but does not include any pseudocode or algorithm blocks.
Open Source Code Yes The code for TVspec NET is publicly available on Github2. 2https://github.com/Tamara Grossmann/TVspec NET
Open Datasets Yes For training and testing our neural network we use the MS COCO dataset [32] that contains a large number of natural images. [32] T.-Y. Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Dollár. Microsoft COCO: Common Objects in Context. ar Xiv preprint ar Xiv:1405.0312v3, 2014.
Dataset Splits No We take 2000 images for training and 1000 for testing. The paper does not explicitly state a separate validation set split or how it was handled (e.g., if a portion of the training set was used for validation).
Hardware Specification Yes on an NVIDIA Quadro P6000 GPU with 24 GB RAM. We also acknowledge the support of NVIDIA Corporation with the donation of two Quadro P6000, a Tesla K40c and a Titan Xp GPU used for this research.
Software Dependencies No The paper mentions software components like Matlab, C++/Python, Adam optimiser, ReLU, Dn CNN, FFDnet, and U-Net, but it does not specify any version numbers for these software dependencies.
Experiment Setup Yes Each image is turned to greyscale and randomly cropped to a 64 64 pixel window. For the purpose of data augmentation, we also take 128 128 crops for some images and downsample them by a factor of 2, obtaining again images of size 64 64. After standardising the dataset to have zero mean and a standard deviation of 1, we generate K = 50 ground truth bands (7) using the model driven approach in Section 2.2. The bands are then combined dyadically to form 6 spectral bands. ... We use the Adam optimiser [27] with an initial learning rate of 10 3 and multi step learning rate decay. Our neural network is trained with a batch size of 8 and for 5000 epochs...